Connectd already does this when we *receive* an error or warning, but
now do it on send. This causes some slight behavior change: we don't
disconnect when we close a channel, for example (our behaviour here
has been inconsistent across versions, depending on the code).
When connectd is told to disconnect, it now does so immediately, and
doesn't wait for subds to drain etc. That simplifies the manual
disconnect case, which now cleans up as it would from any other
disconnection when connectd says it's disconnected.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We want to avoid lost messages in the common cases.
This generalizes our drain code, by giving the subds each 5 seconds to
close themselves, but continue to allow them to send us traffic (if
peer is still connected) and continue to send them traffic.
We continue to send traffic *out* to the peer (if it's still
connected), until all subds are gone. We still have a 5 second timer
to close the connection to peer.
On reconnects, we don't do this "drain period" on reconnects: we kill
immediately.
We fix up one test which was looking for the "disconnect" message
explicitly.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
A subtle case I hadn't come across before: if a child tal_resizes()
its parent while the parent is being deleted, tal gets confused.
The subd destructor does this using tal_arr_remove() on peer->subds,
which is currently being freed:
```
==61056== Invalid read of size 8
==61056== at 0x185632: del_tree (tal.c:417)
==61056== by 0x18560D: del_tree (tal.c:412)
==61056== by 0x185957: tal_free (tal.c:486)
==61056== by 0x1183BC: peer_discard (connectd.c:1861)
==61056== by 0x11869E: recv_req (connectd.c:1942)
==61056== by 0x12774B: handle_read (daemon_conn.c:35)
==61056== by 0x173453: next_plan (io.c:59)
==61056== by 0x17405B: do_plan (io.c:407)
==61056== by 0x17409D: io_ready (io.c:417)
==61056== by 0x176390: io_loop (poll.c:453)
==61056== by 0x118A68: main (connectd.c:2082)
==61056== Address 0x4bd8850 is 16 bytes inside a block of size 48 free'd
==61056== at 0x483DFAF: realloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==61056== by 0x1860E6: tal_resize_ (tal.c:699)
==61056== by 0x1373DD: tal_arr_remove_ (utils.c:184)
==61056== by 0x11D508: destroy_subd (multiplex.c:930)
==61056== by 0x1850A4: notify (tal.c:240)
==61056== by 0x1855BB: del_tree (tal.c:402)
==61056== by 0x18560D: del_tree (tal.c:412)
==61056== by 0x18560D: del_tree (tal.c:412)
==61056== by 0x185957: tal_free (tal.c:486)
==61056== by 0x1183BC: peer_discard (connectd.c:1861)
==61056== by 0x11869E: recv_req (connectd.c:1942)
==61056== by 0x12774B: handle_read (daemon_conn.c:35)
```
So simply make the subds children of `peer` not the `peer->subds`
array. The only effect is that drain_peer() can't simply free the
subds array but must free the subds one at a time.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This allows us to detect when lightningd hasn't seen our latest
disconnect/reconnect; in particular, we would hit the following pattern:
1. lightningd says to connect a subd.
2. connectd disconnects and reconnects.
3. connectd reads message, connects subd.
4. lightningd reads disconnect and reconnect, sends msg to connect to subd again.
5. connectd asserts because subd is alreacy connected.
This way connectd can tell if lightningd is talking about the previous
connection, and ignoere it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Before this patch:
1. connectd says it's connected (peer_connected)
2. we tell connectd we want to talk about each channel (peer_make_active)
3. connectd gives us an fd for each channel, and we connect it to a subd (peer_active)
4. OR, connectd says it sent something about a channel we didn't tell it about, with an fd (peer_active)
Now:
1. connectd says it's connected (peer_connected)
2. we start all appropriate subds and tell connectd to what channels/fds (peer_connect_subd).
3. if connectd says it sent something about a channel we didn't tell it about, we either tell
it to hang up (peer_final_msg), or connect a new opening daemon (peer_connect_subd).
This is the minimal-size patch, which is why we create socket pairs in
so many places to use the existing functions. Many cleanups are
possible, since the new flow is so simple.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Sending any pending messages to peer before hanging up is a courtesy:
give it 5 seconds before simply closing.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Now we have separate peer draining logic, we can simply use it when
connectd tells us to release the peer, without waiting. (We could
simply free the peer, but that's a bit rude, as messages can get
lost).
This removes various complex flags and logic we had before.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Fixed: `connectd`: various crashes and issues fixed by simplification and rewrite.
This removes it from the hashtable, and forces it to do nothing but
send out any remaining packets, then close.
It is, in effect, reduced to a stub, with no further interactions
with the rest of the system (all subds are freed already).
Also removes the need for an explicit "final_msg" too.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This came out in a later patch: freeing the peer->subds doesn't actually
free the subds, because they're reparented onto subd->conn, which is
a child of peer itself.
This breaks because when the peer is finally freed, destroy_subd is
called, and expects to find itself in peer->subds (but we made that
NULL when we manually freed it!).
Fix this, and make it obvious that we tal_steal it.
```
ightning_connectd: FATAL SIGNAL 11 (version v0.11.0.1-25-gbf025aa-modded)
0x55de2a1b8b94 send_backtrace
common/daemon.c:33
0x55de2a1b8c3e crashdump
common/daemon.c:46
0x7fe2be2fc08f ???
/build/glibc-SzIz7B/glibc-2.31/signal/../sysdeps/unix/sysv/linux/x86_64/sigaction.c:0
0x55de2a1af41e destroy_subd
connectd/multiplex.c:1119
0x55de2a217686 notify
ccan/ccan/tal/tal.c:240
0x55de2a217b9d del_tree
ccan/ccan/tal/tal.c:402
0x55de2a217bef del_tree
ccan/ccan/tal/tal.c:412
0x55de2a217bef del_tree
ccan/ccan/tal/tal.c:412
0x55de2a217f39 tal_free
ccan/ccan/tal/tal.c:486
0x55de2a1aa116 peer_discard
connectd/connectd.c:1834
0x55de2a1aa38d recv_req
connectd/connectd.c:1903
0x55de2a1b9121 handle_read
common/daemon_conn.c:31
0x55de2a205a35 next_plan
ccan/ccan/io/io.c:59
0x55de2a20663d do_plan
ccan/ccan/io/io.c:407
0x55de2a20667f io_ready
ccan/ccan/io/io.c:417
0x55de2a208972 io_loop
ccan/ccan/io/poll.c:453
0x55de2a1aa736 main
connectd/connectd.c:2042
0x7fe2be2dd082 __libc_start_main
../csu/libc-start.c:308
0x55de2a1a085d ???
???:0
0xffffffffffffffff ???
???:0
```
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
LND and us send 0xFFFFFFFF to turn off gossip. LDK and Eclair don't
seem to turn off gossip at all, but that's OK.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This is a side-effect of fixing aging: sometimes, we age our
rcvd_filter cache too fast, and thus re-xmit. This breaks
our test, since it used dev-disconnect on the channel_announce,
but that closes to l3, not l1!
```
> assert l1.rpc.listchannels()['channels'] == []
E AssertionError: assert [{'active': T...ags': 1, ...}] == []
E Left contains 2 more items, first extra item: {'active': True, 'amount_msat': 100000000msat, 'base_fee_millisatoshi': 1, 'channel_flags': 0, ...}
```
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Fixes: #5403
Got complaints about us hanging up on some nodes because they don't respond
to pings in a timely manner (e.g. ACINQ?), but that turned out to be something
else.
Nonetheless, we've had reports in the past of LND badly prioritizing gossip
traffic, and thus important messages can get queued behind gossip dumps!
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Changed: connectd: give busy peers more time to respond to pings.
This will be used to decouple internal use of gossip from what is
passed to gossip peers. Updates GOSSIP_STORE_VERION to 10.
Changelog-Changed: gossip_store updated to version 10.
When we moved gossip filtering to connectd, this aging got lost.
Without this, we hit the 10,000 entry limit before expiring full
gossip anti-echo cache. This is under 1M in allocations per peer, but
in DEVELOPER mode each allocation includes adds 3 notifiers (32 bytes
each) and a backtrace child (40 + 40 + 256 bytes), making it almost
10MB per peer, plus allocation overhead.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Fixed: connectd: large memory usage with many peers fixed.
We seem to have made node_announcement propagation *worse*, not
better. Explorers don't see my nodes updates.
At least some LND nodes never send us timestamp_filter, so we are
never actually stream *any* gossip. We should send gossip about
ourselves, even if they haven't set a filter (yet).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Added: Protocol: we more aggressively send our own gossip, to improve propagation chances.
Mostly comments and docs: some places are actually paths, which
I have avoided changing. We may migrate them slowly, particularly
when they're user-visible.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We do this (send warnings) in almost all cases anyway, so mainly this
is a textual update, but there are some changes:
1. Send ERROR not WARNING if they send a malformed commitment secret.
2. Send WARNING not ERROR if they get the shutdown_scriptpubkey wrong (vs upfront)
3. Send WARNING not ERROR if they send a bad shutdown_scriptpubkey (e.g. p2pkh in future)
4. Rename some vars 'err' to 'warn' to make it clear we send a warning.
This means test_option_upfront_shutdown_script can be made reliable, too,
and it now warns and doesn't automatically close channel.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Gossipd didn't actually suppress all gossip, resulting in a flake!
Doing it in connectd now makes much more sense.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
I removed these prematurely: we *haven't* had a release since
introducing them!
This consists of reverting d15d629b8b
"plugins/fetchinvoice: remove obsolete string-based API." and
plugins/fetchinvoice: remove obsolete string-based
API. "onion_messages: remove obs2 support."
Some minor changes due to updated fromwire_tlv API since they
were removed, but not much.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-EXPERIMENTAL: REVERT: Removed backwards compat with onion messages from v0.10.1.
This means doing some wire interpretation, and handling the transient
case where we switch from temporary to permenant channel_id, but it's
not that bad (and required for accurate demux when multiple channels
are involved for a single peer).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Now we always have it (either extracted from an unsolicited message,
or told to us by lightningd when it tells us it wants to talk), we can
always send it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This means lightningd needs to create the temporary one and tell it to
openingd/dualopend, rather than the other way around.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Either because lightningd tells us it wants to talk, or because the peer
says something about a channel.
We also introduce a behavior change: we disconnect after a failed open.
We might want to modify this later, but we it's a side-effect of openingd
not holding onto idle connections.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The message from lightningd simply acknowleges that we are allowed to
discard the peer (because no subdaemons are talking to it anymore).
This difference becomes more stark once connectd holds on to idle
peers.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Use tmpctx, rather than freeing manually everywhere (proof: next patch
added a branch and forgot to free it!).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This happens when we send a warning or lightningd tells us to send a
final message then close. Normally io logging is done by the
subdaemon that creates it, but this is a special case.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This is neater than what we had before, and slightly more general.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Changed: JSON_RPC: `sendcustommsg` now works with any connected peer, even when shutting down a channel.
We don't need to log msgs from subds, but we do our own, and we weren't.
1. Rename queue_peer_msg to inject_peer_msg for clarity, make it do logging
2. In the one place where we're relaying, call msg_queue() directly.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This is critical in the common case where peer sends an error and
hangs up: we almost never get to relay the error to the subd in time.
This also applies in the other direction: we need to flush the queue
to the peer when the subd closes. Note we only free the actual peer
struct when lightningd reaps us with connectd_peer_disconnected().
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We would lose packets sometimes due to this previously, but it
doesn't happen over localhost so our tests didn't notice. However,
now we have connectd being sole thing talking to peers, we can do
a more elegant shutdown, which should fix closing.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Fixed: Protocol: Always flush sockets to increase chance that final message get to peer (esp. error packets).
dev_blackhole_fd was a hack, and doesn't work well now we are async
(it worked for sync comms in per-peer daemons, but now we could sneak
through a read before we get to the next write).
So, make explicit flags and use them. This is much easier now we
have all peer comms in one place.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We actually intercept the gossip_timestamp_filter, so the gossip_store
mechanism inside the per-peer daemon never kicks off for normal connections.
The gossipwith tool doesn't set OPT_GOSSIP_QUERIES, so it gets both, but
that only effects one place.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
