Repository navigation
Description
Issue and Steps to Reproduce
Calling setconfig on a node whose configuration file is not writable by the daemon crashes the entire node with FATAL SIGNAL 6 (abort), instead of returning a JSON-RPC error. The crash happens in the persistence step (setconfig_success → configvar_save → create_setconfig_include → append_to_file), i.e. after the command has been accepted.
Setup (a fairly standard hardened systemd deployment):
- lightningd v26.06.6 started as
lightningd --conf=/etc/lightning/lightning.conf /etc/lightning/lightning.confis owned by root; the service runs asUser=lightningwith systemd hardening, so/etcis read-only to the daemon:[Service] ExecStart=/usr/local/bin/lightningd --conf=/etc/lightning/lightning.conf User=lightning PrivateTmp=true ProtectSystem=full ProtectHome=read-only NoNewPrivileges=true PrivateDevices=true
- Relevant config contents (sanitized):
network=bitcoin lightning-dir=/var/lib/lightning # writable by the daemon rpc-file=/var/lib/lightning-rpc/lightning-rpc rpc-file-mode=0660 - The option being set (
min-capacity-sat) was not previously set in any config file.
Steps:
lightning-cli setconfig min-capacity-sat 2000000
→ returns success to the client, butlistconfigs min-capacity-satstill shows10000/"source": "default", and nothing is persisted. (Silent no-op reported as success — arguably a second bug.)- Run the identical
setconfigagain
→lightning-cligetsreading response: socket closed; lightningd logs**BROKEN** lightningd: Could not write to config : No such file or directoryand aborts.
Note the error message contains an empty filename (to config :), which may point at the underlying cause — the include-file path apparently resolves to an empty string in this layout.
Expected behavior
- If persistence fails,
setconfigshould return a JSON-RPC error (and ideally roll back or keep the value as runtime-only with a warning) — a failedopen()/write()in an RPC handler should neverfatal()the daemon. - A
setconfigcall that changes and persists nothing should not report success.
Actual behavior
The node aborts mid-operation (ours was mid block-scan; systemd's Restart=on-failure brought it back, but on an unsupervised node this is an outage; on a node with active channels it is a crash during live operation, triggerable by any operator using a documented RPC).
Crash log / backtrace
Could not write to config : No such file or directory
2026-07-24T14:42:03.795Z **BROKEN** lightningd: Could not write to config : No such file or directory
lightningd: FATAL SIGNAL 6 (version v26.06.6)
0x600bab42c365 send_backtrace
common/daemon.c:38
0x600bab42c401 crashdump
common/daemon.c:83
0x783308695b2c ???
pthread_kill+0x11c:0
0x78330863c27d ???
gsignal+0x1d:0
0x78330861f8fe ???
abort+0xde:0
0x600bab3bc300 fatal_vfmt
lightningd/log.c:1128
0x600bab3bc39f fatal
lightningd/log.c:1138
0x600bab3ef823 append_to_file
lightningd/configs.c:351
0x600bab3ef99b create_setconfig_include
lightningd/configs.c:496
0x600bab3efc24 configvar_save
lightningd/configs.c:513
0x600bab3f02d8 setconfig_success
lightningd/configs.c:600
0x600bab3f086d json_setconfig
lightningd/configs.c:804
0x600bab3b3c13 command_exec
lightningd/jsonrpc.c:771
0x600bab3b59b1 rpc_command_hook_final
lightningd/jsonrpc.c:912
0x600bab3e6a7d hook_done
lightningd/plugin_hook.c:243
0x600bab3e6bae plugin_hook_call_next
lightningd/plugin_hook.c:343
0x600bab3e78c3 plugin_hook_call_
lightningd/plugin_hook.c:395
0x600bab3b607d plugin_hook_call_rpc_command
A full crash.log from the incident is available on request.
Environment
- Core Lightning v26.06.6 (built from source)
- Debian-based Linux inside a Proxmox LXC container, systemd-managed service
- bitcoind backend on the same host (was still catching up on blocks at the time; likely irrelevant)
Workaround
Avoid setconfig on deployments where the daemon cannot write its own config; edit the config file and restart instead. (Arguably a config file that the daemon cannot rewrite is a deliberate hardening choice, which is why graceful degradation would be valuable here.)
Activity
nepet commentedon Nov 7, 2026
The crash comes from a use-after-free in
setconfig's writability check, introduced in 5e79cd4 (the fix for #7964).With an explicit
--conf,base_conf_file()returnsld->config_filenameitself, and when the config's directory isn't writable,config_not_writable()returns it withtal_steal(cmd, fname)(lightningd/configs.c:644 on master). The firstsetconfigfails with "Cannot write to config file /etc/lightning/lightning.conf", as it should, but freeing that command freesld->config_filename.The second
setconfigthen reads the freed name, which here reads as an empty string.dir_writable("")checks., lightningd's working directory, which is writable, so the check passes, andappend_to_file()aborts on the empty name. That's theCould not write to config : No such file or directorywith no file name in your log.Repro: put the config in a directory lightningd can't write, start with
--conf=<that file>, and callsetconfig min-capacity-sat 20000twice. The first call fails cleanly, the second aborts lightningd.#9711 returns a copy instead. #9356 is still useful as defence in depth, but it doesn't fix this on its own: with it, the second call still reads freed memory.
lightningd: don't free the --conf filename when setconfig can't write it