When we are calling hooks, we track them via a linked list. As they
execute, we pop them off the list in plugin_hook_killed().
When we kill a plugin, we have a destructor which remove its entry from the linked list: plugin_hook_killed.
If it's at the head of the list, that means the plugin died while
processing the hook, so instead of just deleting it, we call
plugin_hook_killed() which behaves as if it said "result: continue".
But plugin_hook_killed() just returns if we're shutting down; this
leaves the link (then freed) on the list, and the *next* plugin tries
to unlink from the list, accessing the previous free entry.
The fix is simple: unlink from the list in plugin_hook_killed() even
if we're shutting down.
```
Valgrind error file: valgrind-errors.78570
==78570== Invalid write of size 8
==78570== at 0x174B55: list_del_ (list.h:328)
==78570== by 0x174FCC: plugin_hook_killed (plugin_hook.c:135)
==78570== by 0x21DC3F: notify (tal.c:240)
==78570== by 0x21E156: del_tree (tal.c:402)
==78570== by 0x21E1A8: del_tree (tal.c:412)
==78570== by 0x21E4F2: tal_free (tal.c:486)
==78570== by 0x16EBD1: plugin_kill (plugin.c:345)
==78570== by 0x16F9C4: plugin_conn_finish (plugin.c:724)
==78570== by 0x20F1A5: destroy_conn (poll.c:244)
==78570== by 0x20F1C9: destroy_conn_close_fd (poll.c:250)
==78570== by 0x21DC3F: notify (tal.c:240)
==78570== by 0x21E156: del_tree (tal.c:402)
==78570== Address 0x6aee688 is 40 bytes inside a block of size 72 free'd
==78570== at 0x483CA3F: free (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==78570== by 0x21E224: del_tree (tal.c:421)
==78570== by 0x21E1A8: del_tree (tal.c:412)
==78570== by 0x21E4F2: tal_free (tal.c:486)
==78570== by 0x16EBD1: plugin_kill (plugin.c:345)
==78570== by 0x16F9C4: plugin_conn_finish (plugin.c:724)
==78570== by 0x20F1A5: destroy_conn (poll.c:244)
==78570== by 0x20F1C9: destroy_conn_close_fd (poll.c:250)
==78570== by 0x21DC3F: notify (tal.c:240)
==78570== by 0x21E156: del_tree (tal.c:402)
==78570== by 0x21E4F2: tal_free (tal.c:486)
==78570== by 0x20D7B6: io_close (io.c:450)
==78570== Block was alloc'd at
==78570== at 0x483B7F3: malloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==78570== by 0x21DCAD: allocate (tal.c:250)
==78570== by 0x21E26E: tal_alloc_ (tal.c:428)
==78570== by 0x175599: plugin_hook_call_ (plugin_hook.c:259)
==78570== by 0x13616F: plugin_hook_call_onion_message_blinded (onion_message.c:126)
==78570== by 0x13643B: handle_obs_onionmsg_to_us (onion_message.c:187)
==78570== by 0x138BBD: gossip_msg (gossip_control.c:140)
==78570== by 0x178AEC: sd_msg_read (subd.c:495)
==78570== by 0x20CA00: next_plan (io.c:59)
==78570== by 0x20D608: do_plan (io.c:407)
==78570== by 0x20D64A: io_ready (io.c:417)
==78570== by 0x20F8F1: io_loop (poll.c:445)
```
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This is required for the setuptools_scm package to correctly identify
the root of the git repository when installing from the
requirements.txt file in the root. It'd otherwise copy the source
directory, which doesn't yet include the version metadata, into a
separate directory and the building from there, which breaks the
git lookup.
`in-tree-build` is the future default, so we'll eventually be able to
strip that option again. See [1] for details.
[1] https://github.com/pypa/pip/issues/7555
When downloading a python package from the PyPI repository the links
where pointing to a non-existent parent directory, thus breaking the
packages. The files don't ever change, and are really simple, so let's
just materialize them.
It was incredibly flaky due to the potential for l2 announcing the
channel before l1 could get to it, thus suppressing the outgoing
announcement which we were looking for. This now checks either
direction.
Before this fix the failure rate was 24% (out of 100 runs), afterwards
it's 0%.
Changelog-None
This was measured as a 95th percentile in our rough testing, thanks to
all the volunteers who monitored my channels.
Fixes: #4761
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Added: JSON-RPC: `setchannelfee` gives a grace period (`enforcedelay`) before rejecting old-fee payments: default 10 minutes.
We can miss it in both logs, so wait for it instead:
```
2021-09-22T07:25:59.1582950Z > l3.rpc.addgossip(ann.split()[3])
2021-09-22T07:25:59.1583911Z E AttributeError: 'NoneType' object has no attribute 'split'
```
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
By popular merge-hell demand.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Changed: Build: Python is now required to build, as generated files are no longer checked into the repository.
Suggested by @cdecker
P.S: Also this include an API refactoring from my previous solution, also this it is suggested by @cdecker.
Signed-off-by: Vincenzo Palazzo <vincenzopalazzodev@gmail.com>
Changelog-Added: Support to listpays the status parameter to filter the payments by status.
Signed-off-by: Vincenzo Palazzo <vincenzopalazzodev@gmail.com>
Currently it will be used for onion replies, but we can use it for offers
and invoices in future, if we want to avoid revealing our node_id.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This expects the caller to create the TLVs to put in each hop; it
simply creates the onion and sends it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We put this in reply paths, so we can tell if they are used. This lets us
avoid responding unless the correct reply path is used.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
One change from the obsolete version handling, gossipd will no longer send
forwarding onion msgs to lightningd, but will forward it directly.
That was the effect before, anyway.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
It was really different from the way we decide the overall state of a
`pay` command's output. Now we use a more similar state decision,
based on collecting all states and checking them at the end to
determine the outcome.
We were checking against the wrong enum value, resulting in an
erroneous `pending` state as a final result.
Changelog-Fixed: pay: `pay` would sometimes misreport a final state of `pending` instead of `failed`
We had sent->path be a list of node_ids, but it makes more sense as
pubkeys so we can avoid conversion. Also, I find it easier to think
about (especially creating backwards paths) if we include *ourselves*
as the first element in the path.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
