Currently, if we don't realize a TCP connection is down, we almost
certainly don't find out until *after* we're sent the
commitment_signed message, in which case we cannot fail the incoming
HTLC.
This test demonstrates that. Note the 30 second sleep: we should really
run Travis tests in parallel!
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Avoid that 200ms loss. We don't want to disable nagle generally,
since it's great for gossip and other traffic; we just want to push at
critical times.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Since we now fixed the bug where nodes receiving a connection would
try to reconnect to the source IP/port of that connection, we now expose
an issue mentioned by other implementers: we can continually cross over
reconnections unless we add some fuzz. One second should be sufficient.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If we have an address hint, we start with that, but we'll use
node_announcement information if required.
Note: we (ab)use the address hint when restoring from the database
or reconnecting, even if the connection was *incoming*. That meant
that the recipient of a connection would *never* manage to connect out.
We still don't take multiple addresses from the DNS seeds: I assume we
should, since there could be IPv4 and IPv6.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We're currently overriding fatal() with something that actually
returns, which contrasts with its declaration as NORETURN.
This breaks in the next patch which wants a real fatal() in wallet.h.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
[ Squashed into single commit --RR ]
This adds two new macros, `p_req_tal()` and `p_opt_tal()`. These support
callbacks that take a `struct command *` context. Example:
static bool json_tok_label_x(struct command *cmd,
const char *name,
const char *buffer,
const jsmntok_t *tok,
struct json_escaped **label)
The above is taken from the run-param unit test (near the bottom of the diff).
The return value is true on success, or false (and it calls command_fail itself).
We can pretty much remove all remaining usage of `json_tok_tok` in the codebase
with this type of callback.
So much code for so little noticable difference, but mention that
`state` (an accidental legacy from before we moved the field into
`channels` long ago) no longer exists.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We currently hand the error back to the master, who then stores it for
future connections and hands it back to another openingd to send and exit.
Just send directly; it's more reliable and simpler.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Include it as an optional field in the connect_to_peer message (it was
added before we had optional fields).
The only issue is that reconnects want it too, so again connectd hands
it back to master in connectctl_connect_failed.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
connectd tells master about every disconnection, and master knows
whether it's important to reconnect. Just get the master to invoke a new
connect command if it considers the peer important!
The only twist is timeouts: we don't want to immediately reconnect if
we've failed to connect. To solve this, connectd passes a 'delaytime'
to the master when a connection fails, and the master passes it back
when it asks for a connection.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We used to separate implicit connection requests (ie. timed retries
for important peers) and explicit ones, and send a
WIRE_CONNECTCTL_CONNECT_TO_PEER_RESULT for the latter.
In the success case, that's now redundant, since we hand the connected
peer to the master using WIRE_CONNECT_PEER_CONNECTED; we just need a
message for the failure case. And we might as well tell the master
every failure, so we don't have to distinguish internally.
This also solves a race we had before: connectd would send
WIRE_CONNECTCTL_CONNECT_TO_PEER_RESULT which completes the incoming
JSON connect command, then send WIRE_CONNECT_PEER_CONNECTED. So
there's a window where the JSON command can return, but the peer isn't
known to lightningd yet.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
In particular, all opening_read_peer_msg() callers need to know there
was an error (presumably, negotiating) so they can stop, but we should
not exit.
This lets us reenable the final disabled test.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We don't want to exit just because channel parameter negotiation
failed, but we do want to tell the master if it was a channel we were
trying to fund.
Note that lightningd still needs to fail the funding cmd if it gets a
fromwire_opening_fundee (they raced us and won), or an outright
failure.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Previously master would fail once the channel has been negotiated,
which is terrible, since the funder will have already broadcast tx.
Now we tell them if we have an active channel, and update if it goes away.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We now simply maintain a pubkey set for connected peers (we only care
if there's a reconnect), not the entire peer structure.
lightningd no longer queries us for getpeers: it knows more than we do
already.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
There's now a potential race: the source peer connect returns, but in
destination peer the master hasn't read the connect message from
connectd, so the peer isn't in listpeers yet.
(Previously the connection stayed in connectd, so there was no such
window).
This is an occasional issue in a few places.
Note that we take the opportunity to speed up test_disconnectpeer too
while we're there.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Prior to this, lightningd would hand uninteresting peers back to connectd,
which would then return it to lightningd if it sent a non-gossip msg,
or if lightningd asked it to release the peer.
Now connectd hands the peer to lightningd once we've done the init
handshake, which hands it off to openingd.
This is a deep structural change, so we do the minimum here and cleanup
in the following patches.
Lightningd:
1. Remove peer_nongossip handling from connect_control and peer_control.
2. Remove list of outstanding fundchannel command; it was only needed to
find the race between us asking connectd to release the peer and it
reconnecting.
3. We can no longer tell if the remote end has started trying to fund a
channel (until it has succeeded): it's very transitory anyway so not
worth fixing.
4. We now always have a struct peer, and allocate an uncommitted_channel
for it, though it may never be used if neither end funds a channel.
5. We start funding on messages for openingd: we can get a funder_reply
or a fundee, or an error in response to our request to fund a channel.
so we handle all of them.
6. A new peer_start_openingd() is called after connectd hands us a peer.
7. json_fund_channel just looks through local peers; there are none
hidden in connectd any more.
8. We sometimes start a new openingd just to send an error message.
Openingd:
1. We always have information we need to accept them funding a channel (in
the init message).
2. We have to listen for three fds: peer, gossip and master, so we opencode
the poll.
3. We have an explicit message to start trying to fund a channel.
4. We can be told to send a message in our init message.
Testing:
1. We don't handle some things gracefully yet, so two tests are disabled.
2. 'hand_back_peer .*: now local again' from connectd is no longer a message,
openingd says 'Handed peer, entering loop' once its managing it.
3. peer['state'] used to be set to 'GOSSIPING' (otherwise this field doesn't
exist; 'state' is now per-channel. It doesn't exist at all now.
4. Some tests now need to turn on IO logging in openingd, not connectd.
5. There's a gap between connecting on one node and having connectd on
the peer hand over the connection to openingd. Our tests sometimes
checked getpeers() on the peer, and didn't see anything, so line_graph
needed updating.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Useful for debugging: it wasn't immediately obvious from the logs
which side was spuriously reconnecting.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
In particular, I found lightning_openingd processes after running
tests. When we use the dev_disconnect blackhole '0' option, they
stick around until the dev_disconnect file is truncated (there is only
so much you can do with only a file descriptor), so let's do that.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The following changes revealed this race, where expecting listchannels()
to contain two channels immediately after fund_channel() was racy.
We also derive the short_channel_id first, so we can search logs for the
exact messages.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The next patches get better at reconecting, so if we use dev-allow-localhost
nodes can often find each other and reconnect before shutting down; only
use that option where we actually need it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Saw this in Travis: technically we return from the dev_set_max_scids...
cmd after sending it to gossipd, but we should wait for it to log.
Adding an internal reply message for a dev command seems overkill.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
They're much more useful being programatically-accessible, AFAICT.
The string stays the same so they're backwards compatible.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1. If the IPv6 address was public, that changed the wireaddr and thus the ipv4 bind
would not be to a wildcard and would fail.
2. Binding two fds to the same port on both wildcard IPv4 and IPv6 succeeds; we only
fail when we try to listen, so allow error at this point.
For some reason this triggered on my digital ocean machine.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>