This was a regression from e83f0dd3fc,
introduced by the change in connection_down() there.
(if the initial connection to the main interface is not successful, the
network status will not be set to 'disconnected' - but it should be)
That change is now reverted here.
That change was somewhat independent of the rest of that commit,
except the rest highlighted that something of the sort was needed:
as it might sometimes take many seconds for an interface to close
and we might launch another interface for the same server while the
first one is still closing, the "server == self.default_server" test
is highly problematic.
To fix this second - original - issue, we now introduce a "closing" state
for interfaces (in the form of network._closing_ifaces).
The GUI blocks until network.set_parameters returns when switching servers,
which waits for switch_to_interface, which used to wait until interface.close()
returns. interface.close() tries to flush buffered writes to the wire, with a
30 sec timeout.
If the server or the network connection is slow, flushing the buffer can take
several seconds. In particular, servers running Fulcrum always seem to
timeout in this case, freezing the GUI for 30 seconds (when switching away).
The gossip db is loaded early when the network is started to save
time when the gui is locked and a wallet not yet loaded. Side effects
of the LNWallet to start peering when a channel db is loaded is
circumvented.
Enables lightning by creating a node private key and storing it in
the wallet. The gossiper is not launched at start up, only if there
are existing channels.
some basic sanity checks
Previously if the server sent back a malformed response, it could partially corrupt a wallet file.
(as sometimes the response would get persisted, and issues would only arise later when the values were used)
Previously this function would not switch to a different chain if the
current chain contained the preferred block. This was not the intended
behaviour: if there is a *stronger* chain that *also* contains the
preferred block, we should jump to that.
Note that with this commit there will now always be a preferred block
(defaults to genesis). Previously, it might seem that often there was none,
but actually in practice if the user used the GUI context menu to switch
servers even once, there was one (usually genesis).
Hence, with the old code, if an attacker mined a single header which
then got reorged, auto_connect clients which were connected to the
attacker's server would never switch servers (jump chains) even
without the user explicitly configuring preference for the stale branch.
Data returned by the main server for request "server.peers.subscribe"
is of course untrusted input. Previously if it contained e.g. invalid port numbers
or IP addresses, it could kill the whole network taskgroup.
(this might have only affected master and not released versions,
which would only raise exceptions once the client actually tried to connect to an invalid host/port)
Previously all the interfaces used either "t" or "s".
Now the network only tries to use "s" for all interfaces, except for
the main interface, which the user can manually specify to use "t".
(so e.g. if you run with "--server localhost:50002:t", the main server will use "t",
but all the rest will use "s")
network.main_taskgroup restarts every time the proxy settings are changed,
many long-running tasks (some introduced with lightning) are not prepared for and do not want this.
