We allocate the default, then callback allocates over the top. Mark
params with a default, so we can free that when it's called.
(We can't do this generally, since not all param args are actually
pointers to pointers, though opt_param_def has to be).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1. Tell memleak about our linked-list of current payments.
2. Don't remove them from list until we actually free them (in destructor, naturally).
3. Decode invoices into tmpctx (we steal / copy what we want anyway).
4. Free params after we've used them.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1. The dijkstra can be temporary, doesn't need to last as long as pay cmd.
2. We fail multiple times in several places, so don't leak old failreason.
3. Make payments findable by our memleak detector.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1. We don't keep a pointer to payments (unlike pay, which keeps a
linked list), so mark it notleak.
2. plugin->our_features is overloaded for "features we want to set" (used by keysend)
and then "features we have". Create a new field, which is cleaner.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
1. p is a child of cmd, so it's freed by command_failed.
2. cltv_budget is set a few lines up to the same thing already.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Seeing if this helps my build box which previously gives:
```
/home/rusty/lightning-ltest/lightningd/lightning_connectd: libbacktrace: unrecognized DWARF version in .debug_info at 6
/home/rusty/lightning-ltest/lightningd/lightning_connectd: libbacktrace: no debug info in ELF executable
```
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
These tests have proven to be:
a) very expensive, as they spin up many nodes, and perform long setup
b) are not testing anything specific, they just fuzz functionality
that is already tested otherwise
c) have not helped pinpoint any issues in living memory
d) are very flaky, making for really bad signal-to-noise, so much
that devs usually just restart without even looking at the logs
e) even if we were to look at the logs, we'd be unable to reproduce
due to the inherent randomness involved in these tests
f) are really noisy neighbors, causing other tests to flake as well,
further muddying the water
All in all, these tests are a waste of time, and source of
frustration.
[ Cleaned up python unused imports --RR ]
Changelog-None
client_read_next(…) calls io_read_wire(…), passing &c->msg_in as the
address of a pointer that will be set to the address of the buffer that
io_read_wire_(…) will allocate, and passing c (a pointer to the struct
client instance) as the parent for the new allocation. As long as the
struct client instance eventually gets freed, the allocated message
buffer will be freed too, so there is no "leak" in the strict sense of
the term, but the freeing of the buffer may not occur for an arbitrarily
long time after the buffer has become disused, and indeed many millions
of message buffers may be allocated within the lifetime of one struct
client instance.
handle_client(…) ultimately hands off the c->msg_in to one of several
message-type-specific handler functions, and those functions are not
TAKES or STEALS on their message buffer parameters and do not free their
message buffer arguments. Consequently, each successive call to
client_read_next(…) will cause io_read_wire_(…) to overwrite the
c->msg_in pointer with the address of a newly allocated message buffer,
and the old buffer will be left dangling off of the struct client
instance indefinitely.
Fix this by initializing c->msg_in to NULL in new_client(…) and then
having client_read_next(…) do `c->msg_in = tal_free(c->msg_in)` prior to
calling io_read_wire(…). That way, the previous message buffer will be
freed just before beginning to read the next message. The same strategy
is already employed in common/daemon_conn.c, albeit without nulling out
dc->msg_in after freeing it.
Fixes: #5035
Changelog-Fixed: hsmd: Fixed a significant memory leak
This restores the behaviour prior to `lightningd: use our cached
channel_update for errors instead of asking gossipd.`, where gossipd
would refuse to give us channel_updates for unannounced channels.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
It looks like decode_c doesn't set have_c unlike the other decode_
methods. At the start of the function, decode_c checks have_c to see if
it's set, but it is never set. It seems like this could allow for
duplicate c tags, which is probably not intended.
Signed-off-by: William Casarin <jb55@jb55.com>
If this field is missing for whatever reason (weird db state?)
clightning will crash when listing invoices.
Signed-off-by: William Casarin <jb55@jb55.com>
Unfortunately we can't do any smart parsing here since
wiregen does not support switch/type cases for different
substructure unions yet. So just give us a pointer we can use.
But this requires a watch-only wallet, and python-bitcoinlib doesn't support
multiple wallets, so we need to unload the original one, but then we need
to generate a block, so that can't generate a new address, so we need
an address arg to generate_block.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We detect whether we have the rust tooling available (mainly `cargo`)
and enable or disable the rust libraries, plugins and examples when it
is enabled. Since the rest of the Makefiles assumes that executables
have an associated header and C source file, we also needed to add a
target that we can add non-C binaries to.
We build an in-memory model of what the API should look like, which
will later be used to generate a variety of bindings. In this PR we
will use the model to build structs corresponding to the requests and
responses for the various methods.
The JSON-RPC schemas serve as ground-truth, however they are missing a
bit of context: methods, and the request-response matching (as well as
a higher level grouping we'll call a Service). I'm tempted to create a
new document that describes this behavior and we could even generate
the rather repetitive JSON schemas from that document. Furthermore
it'd allow us to add some required metadata such as grpc field
numbering once we generate those bindings.
Changelog-Added: JSON-RPC: A new `msggen` library allows easy generation of language bindings for the JSON-RPC from the JSON schemas
I've tried automatically parsing the docs, and these inconsistencies made it harder to do that.
(I tried to do that for a project which I can't share yet, I'm not sure if it'll even work).
We really need our own lnprototest tests for packet-based stuff;
these message-based tests are inherently delicate and awkward.
In particular, connectd now does dev-disconnect, so the socket is not
immediately closed after a dev-disconnect command. In this case, the
WIRE_SHUTDOWN has often already been written from connectd to channeld.
But it sometimes works, too.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
If the HTLCs are completely negotiated, we can get a channel break when
we mine a pile of blocks. This is mainly seen with Postgres, due to the db
speed.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
