On Tue, Jul 16, 2019 at 10:24:24PM -0500, Eric Blake wrote:
+static gboolean +prepare (GSource *sp, gint *timeout_) +{ + struct NBDSource *source = (struct NBDSource *) sp; + + /* When the NBD handle moves out of the created state (which means + * that it first has a socket associated with it) we must initialize + * and register the pollfd. + */ + if (!source->poll_registered && !nbd_aio_is_created (source->nbd)) { + int fd; + + if (source->connecting_callback) { + DEBUG (source, "calling connecting_callback"); + source->connecting_callback (source); + }
If I understand nbd_aio_connect_tcp properly, there are cases where we end up trying several options presented from getaddrinfo() with a possible reset back to START, at least until we get an fd that sticks. Does this need to take into account that an fd might change while still trying to do the initial connect?
Urgh yes, this is true. There's a second problem which is that if the connection becomes dead then the poll fd becomes invalid. I think all this can be solved but it involves some major reworking. Still working on it ...
+ /* Run the main loop until quit. */ + g_main_loop_run (loop); + exit (EXIT_SUCCESS);
I guess the condition for causing the main loop to run is later on in one of the callbacks?
I think you meant to say "to quit"? Yes, when we detect that the copy has finished we call g_main_loop_quit in one of the callbacks.
+/* This callback is called from libnbd when any read command finishes. */ +static int +finished_read (void *vp, int64_t cookie, int *error) +{ + size_t i; + + if (gssrc == NULL) + return 0; + + DEBUG (gssrc, "finished_read: read completed"); + + /* Find the corresponding buffer and mark it as completed. */ + for (i = 0; i < nr_buffers; ++i) { + if (buffers[i].cookie == cookie) + goto found; + } + /* This should never happen. */ + abort (); + + found: + buffers[i].state = BUFFER_READ_COMPLETED;
You asked on IRC if we should change the semantics of the *_callback variants to auto-retire the command (right now, the commands are not retired until nbd_aio_command_completed, but you can't call that from this callback because of the nbd handle lock - and calling it shouldn't reveal any more information than what you already have access to here). That may be a good idea to play with.
Yes, this is a problem still. But at least I worked out why callbacks were (appearing to be) lost. They aren't, but I was using the cookie to find the command in the list of buffers. Unfortunately I forgot that cookies are not unique between libnbd handles ... Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-builder quickly builds VMs from scratch http://libguestfs.org/virt-builder.1.html