On Tue, Nov 11, 2025 at 04:48:04PM +0100, Kevin Wolf wrote: > Am 11.11.2025 um 15:43 hat Daniel P. Berrangé geschrieben: > > On Wed, Nov 05, 2025 at 03:57:29PM -0600, Eric Blake wrote: > > > On Tue, Nov 04, 2025 at 11:13:48AM +0000, Daniel P. Berrangé wrote: > > > > On Mon, Nov 03, 2025 at 02:10:57PM -0600, Eric Blake wrote: > > > > > The point of QIONetListener is to allow a server to listen to more > > > > > than one socket address at a time, and respond to clients in a > > > > > first-come-first-serve order across any of those addresses. While > > > > > some servers (like NBD) really do want to serve multiple simultaneous > > > > > clients, many other servers only care about the first client to > > > > > connect, and will immediately deregister the callback, possibly by > > > > > dropping their reference to the QIONetListener. The existing code > > > > > ensures that all other pending callbacks remain safe once the first > > > > > callback drops the listener, by adding an extra reference to the > > > > > listener for each GSource created, where those references pair to the > > > > > eventual teardown of each GSource after a given callbacks has been > > > > > serviced or aborted. But it is equally acceptable to hoist the > > > > > reference to the listener outside the loop - as long as there is a > > > > > callback function registered, it is sufficient to have a single > > > > > reference live for the entire array of sioc, rather than one reference > > > > > per sioc in the array. > > > > > > > > > > Hoisting the reference like this will make it easier for an upcoming > > > > > patch to still ensure the listener cannot be prematurely garbage > > > > > collected during the user's callback, even when the callback no longer > > > > > uses a per-sioc GSource. > > > > > > > > It isn't quite this simple. Glib reference counts the callback > > > > func / data, holding a reference when dispatching the callback. > > > > > > > > IOW, even if the GSource is unrefed, the callback 'notify' > > > > function won't be called if the main loop is in the process > > > > of dispatching.
snip > > That appears to work ok, however, there's still a race window that is > > not solved. Between the time thread 2 sees POLLIN, and when it calls > > the dispatch(sock) function, it is possible that thread 1 will drop > > the last reference: > > > > > > > > Thread 1: > > qio_net_listener_set_client_func(lstnr, f, ...); > > => object_ref(listener) > > => foreach sock: socket > > => sock_src = qio_channel_socket_add_watch_source(sock, > > ...., lstnr, NULL); > > > > Thread 2: > > poll() > > => event POLLIN on socket > > > > Thread 1: > > qio_net_listener_set_client_func(lstnr, NULL, ...); > > => foreach sock: socket > > => g_source_unref(sock_src) > > => object_unref(listener) > > unref(lstnr) (still 1 reference left) > > Is what you're worried about that there is still a reference left in > the opaque pointer of an fd handler, but it's not in the refcount and > therefore this already frees the listener while thread 2 will still > access it? Yes, exactly. > > > > > Thread 2: > > => call dispatch(sock) > > => ref(lstnr) > > ...do stuff.. > > => unref(lstnr) (the final reference) > > => finalize(lstnr) > > => return dispatch(sock) > > => unref(GSourceCallback) > > > > > > I don't see a way to solve this without synchronization with the event > > loop for releasing the reference on the opaque data for the dispatcher > > callback. That's what the current code does, but I'm seeing no way for > > the AioContext event loop callbacks to have anything equivalent. This > > feels like a gap in the AioContext design. > > I think the way you would normally do this is schedule a BH in thread 2 > to do the critical work. If you delete the fd handler and unref the > listener in thread 2, then there is no race. Yes, using a BH would be safe, provided you put the BH in the right loop, given that we have choice of the main event loop, or a non-default GMainContext, or the special AioContext that NBD is relying on. > But maybe adding a callback for node deletion in AioHandler wouldn't > hurt because the opaque pointer pretty much always references something > and doing an unref when deleting the AioHandler should be a pretty > common pattern. That would likely make the scenario less error-prone, compared to remembering to use a BH to synchronize. > > This is admittedly an incredibly hard to trigger race condition. It would > > need a client to be calling a QMP command that tears down the NBD server, > > at the exact same time as a new NBD client was incoming. Or the same kind > > of scenario for other pieces of QEMU code using QIONetListener. This still > > makes me worried though, as rare races have a habit of hitting QEMU > > eventually. > > Aren't both QMP and incoming NBD connections always handled in the main > thread? I'm not sure if I know a case where we would actually get this > pattern with two different threads today. Of course, that doesn't mean > that we couldn't get it in the future. Yeah, I believe we're probably safe in todays usage. My concern was primarily about any surprises in the conceptual design that might impact us in future. I guess if NBD is the only thing using AioContext for QIONetListener today, we could hoist the ref/unref only when using a AioContext, and keep the GDestroyNotify usage for the GMainContext code path to significantly limit the exposure. Avoid needing to do anything extra for AioHandlers right before release. With regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
