Previously, onionAddresses definitions in onionServices were of the form
onionAddresses.access.<service> = [<service>];
This caused failures for configurations where a service user name was
overridden or for bitcoind whose default user is 'bitcoin' instead of 'bitcoind'.
Now set the equivalent of:
onionAddresses.access.<actualServiceUser> = [<service>];
Implement this via a new option `onionAddresses.services` to make things more
readable and to work around an infinite recursion error in onionServices.
- enable usage outside of secure-node.nix
- use json as the output format
- show ports
- also show local addresses, which is particularly useful when
netns-isolation is enabled
- only show enabled services
Move 'enforceTor' and onion-service definitions from secure-node.nix.
Use the onionServices module to define onion services.
Onion services now automatically work for services that bind to an INADDR_ANY (`0.0.0.0`) address.
- use -e to check for existence of /var/lib/tor/state, use shorter
polling interval
- clear existing dataDir contents to avoid accumulating obsolete data
- use concatMapStrings instead of foldl'
The 'scudo' memory allocator set by the 'hardened' profile breaks some
services on 20.09.
The fix for NixOS unstable (https://github.com/NixOS/nixpkgs/pull/104052)
is ineffective on 20.09.
As a workaround, add a custom 'hardened' preset that uses the default allocator.
Under normal circumstances, service-specific netns should never exist
before the netns setup service starts.
An existing netns is a genuine error that should not be silently ignored.
A short time after `netns delete` finishes, the peer link in the main
netns is automatically removed.
When `link del` is run before that, it fails with
`Cannot find device "nb-veth-br-*"` and the netns service enters a failed state.
This also enables the test scenario 'netnsRegtest' introduced in a
later commit by fixing the following bug:
For unknown reasons, when tor.active=true and tor is not running, lnd
fails with a tor connection error on netns-isolation, but runs fine
without netns-isolation.
- Remove redundant option mainchainrpchost.
This option is already provided by bitcoind.
- Set a working default for rpcport and rpcuser.
Enables use without secure-node.
a89a3e934f test: increase diskSize (nixbitcoin)
24b506ff8a tests: simplify lightning-loop test (nixbitcoin)
e7c5f956ea lightning-loop: update module (nixbitcoin)
4a503f57bd lightning-loop: v0.8.1 -> v0.9.0 (nixbitcoin)
Pull request description:
ACKs for top commit:
jonasnick:
reACK a89a3e934f
erikarvstedt:
I think it's okay if you would just merge 24b506ff8a, which is the direct parent of the ACK'd a89a3e934f, and removing a89a3e934f itself is totally uncontroversial.
Tree-SHA512: cee2a2714c714a22c35cea0fa829b42a371540983609cda6609f4d063d849f2e725643bd77cfe78eb71665725164d63f83b6c2589be9e72ba30aaecd7c8dee6c
Because we have btcpayserver now, nanopos is not really needed any more. Nanopos
was meant to be just a PoC. Lightning charge can be removed because nanopos is
the only module that depends on it.
- Explain why we don't use option `networking.bridges`
- Make the bridge setup service part of NixOS' network-setup.service.
This yields no noticable functional changes for now, but it's
conceptually cleaner to finish the network setup before network.target
becomes active.
- Add 'nb-' prefix to service name
Going without the conditionals (like in secure-node.nix) adds
readability and doesn't reduce evaluation performance (in fact, it
even slightly improves performance due to implementation details
of mkIf).
To avoid errors, remove use of disabled services in secure-node.nix and
nix-bitcoin-webindex.nix.
The 'availableNetns' connection matrix only consists of enabled entries,
so no extra filtering is needed.
Reason: availableNetns starts with the filtered 'base' and is then symmetrised.
Previously, systemd-status was broken for all users except root.
Use a 'default' deny policy, which is overridden for group 'proc'.
Add operator to group 'proc'.
Also, remove redundant XML boilerplate.
Previously, due to the dependency on a helper service, this dbus config
was initially inactive after system boot, allowing for unrestricted use
of the problematic dbus call.
This also broke the accompanying VM test on faster systems.
Remove 'allow' policy for root because it's a no-op:
1. It's overridden by the 'mandatory' deny policy.
2. Root can use all dbus calls anyways, regardless of policy settings.
Also, add some comments.
Mitigates a security issue that allows unprivileged users to read other
unprivileged user's processes' credentials from CGroup using `systemctl
status`.
- Adds recurring-donations to netns-isolation.services
- Adds cfg.enforceTor to bring recurring-donations in line with other
services
- Removes torsocks dependency in favor of `curl --socks-hostname`
- Adds nanopos to netns-isolation.services
- Adds cfg.enforceTor and extraArgs to bring nanopos in line with other
services
- Adds charged-url option to allow using nanopos with network
namespaces.
- Modularizes nginx so webindex can be used without nanopos.
- Adds host option (defaults to localhost) as target of hidden service
- Removes unnecessary after
- Adds lightning-charge to netns-isolation.services
- Adds cfg.enforceTor to bring lightning-charge in line with other
services
- Adds extraArgs option to allow using lightning-charge with network
namespaces
- Adds host option (defaults to localhost) as target of hidden service
- Adds spark-wallet to netns-isolation.services
- Adds extraArgs option to allow using spark-wallet with network
namespaces
- Adds host option (defaults to localhost) as target of hidden service
- Adds enforceTor option to bring in line with other services
- Adds electrs to netns-isolation.services
- Adds daemonrpc option and specifies address option to allow using
electrs with network namespaces
- Adds host option (defaults to localhost) as target of hidden service
- Adds liquidd to netns-isolation.services
- Adds rpcbind, rpcallowip, and mainchainrpchost options to allow using
liquidd with network namespaces
- Adds bind option (defaults to localhost) as target of hidden service
- Adds lnd to netns-isolation.services
- Specifies listen option (defaults to localhost) as target of
hiddenService.
- Amends hardcoded lnd ip to lnd-cert
WARNING: Breaking changes for lnd cert. lnd-key and lnd-cert will have
to be deleted and redeployed.
- Adds bitcoind-host, and tor-socks options to allow using with
network namespaces.
- Adds listen, rpclisten, and restlisten option to specify host on which
to listen on for peer, rpc and rest connections respectively
- Adds announce-tor option and generates Tor Hidden Service with nix
instead of lnd to bring in line with clightning.
WARNING: Breaking changes for Tor Hidden Service. Manual migration
necessary.
- Adds clightning to netns-isolation.services
- Adds bitcoin-rpcconnect option to allow using clightning with network
namespaces
- Uses bind-addr option (defaults to localhost) as target of hidden service
- Adds different bind-addr options depending on if netns-isolation is
enabled or not.
From the clightning manpage:
autolisten=BOOL By default, we bind (and maybe announce) on IPv4 and
IPv6 interfaces if no addr, bind-addr or announce-addr options are
specified. Setting this to false disables that.
We already set bind-addr by default, so autolisten had no effect.
Therefore, this commit replaces autolisten with the more granular
announce-addr option.
For now we are Tor-only, so we only need to announce our hidden service
to accept incoming connections. In the future, we can add clearnet
connectivity with `addr` and route connections into our netns with NAT.
- Adds bitcoind to netns-isolation.services
- Adds rpcbind and rpcallowip options to allow using bitcoind with
network namespaces
- Adds bind option (defaults to localhost), used as target of hidden service
- Makes bitcoind-import-banlist run in netns
- Adds network namespace instantiation and routing architecture.
- netns-isolation disabled by default. Can be enabled with
configuration.nix FIXME.
- Uses mkMerge to toggle certain options for non netns and netns
systems.
- Adds security wrapper for netns-exec which allows operator to exec
with cap_sys_admin
- User can select the 169.254.N.0/24 addressblock netns's are created in.
- nix-bitcoin-services IpAddressAllow is amended with link-local
addresses
Remove the possible null value for bitcoind.prune and set prune = 0 in
bitcoind as a default. Remove prune = 0 in secure-node.nix and the
mkForce in configuration.nix (bitcoind.prune = lib.mkForce ).
db48ab9b69 services: use 'port' option type (Erik Arvstedt)
Pull request description:
ACKs for top commit:
jonasnick:
ACK db48ab9b69
Tree-SHA512: 24cf0c307b40652d1275575fdf4216696890b0f7786832e7bbee9e21cf6d23d3fc35480926c475fc98c17eba668f5ee2c8c0875689e725c8ad05f2fb6b9ecd20
Add ProtectSystem=strict, remove PermissionStartOnly.
Extract the section of postStart that needs secrets dir write
access into a separate script with full privileges.
Simplify preStart and fix dataDir quoting.
Remove PermissionsStartOnly for bitcoind and spark-wallet (it was never
needed there)
Give reason for PermissionsStartOnly in lightning-charge
Replace PermissionsStartOnly in clightning, electrs and liquid
This is NixOS' recommended way to setup service dirs
https://github.com/NixOS/nixpkgs/pull/56265. This commit hands off the
initial data directory creation to systemd.tmpfiles.rules. All other
preStart scripts are left intact to limit this changes' scope.
In conjuction with secure-node.nix, this sets sane
RestrictAddressFamilies unless lnd is enabled. Before, we were
constantly exposing unnecessary Address Families, not just when lnd is
enabled.
However, zmqpub* must always be enabled for lnd, even when used
outside of secure-node.nix, so we make this change in the lnd module.
currently, nodeinfo has presets/secure-node.nix as a strict
dependency as it requires onion-chef and the 'operatorName' option.
and nix-bitcoin-webindex.nix has nodeinfo as a dependecy.
so don't add nodeinfo and webindex to modules.nix because they will fail on standalone use.
Electrs does not need to be a part of "bitcoinrpc" group because preStart
electrs.toml creation is handled by PermissionsStartOnly. "bitcoin"
group membership is only necessary when cfg.high-memory is enabled and
electrs reads blocks directly from the blocks directory.
It doesn't make sense for bitcoind users to completely redefine their
config file. Also, it's poorly named and the description is faulty.
This is a breaking change, but this option has probably no actual users.
This fixes these flaws in `copy-root-authorized-keys`:
- When `.vbox-nixops-client-key` is missing, operator's authorized_keys
file is always appended to, growing the file indefinitely.
- Service is always added and not restricted to nixops-vbox deployments.
Electrs allows defining settings multiple times via cmdline args, but
not via config files.
So 'extraArgs' is the only way to implement overridable settings,
'extraOptions' wouldn't work.
ProtectSystem=full disables writing to /etc which is the default
secrets location.
Besides that, hardening is pointless for {generate,setup}-secrets which
don't read external input and are fully under our control.
187ff884db add modules test (Erik Arvstedt)
826245484e make secrets dir location configurable (Erik Arvstedt)
b1e13e9415 simplify secrets file format (Erik Arvstedt)
314272a228 lnd, nanopos: move user and group definitions to the bottom (Erik Arvstedt)
766fa4f300 travis: cache all build outputs with cachix (Erik Arvstedt)
b0e759160d travis: set NIX_PATH as early as possible (Erik Arvstedt)
c51bbcf104 travis: move comment (Erik Arvstedt)
7092dce0c7 travis: remove use of deprecated statements (Erik Arvstedt)
190a92507c travis: split up scripts into statements (Erik Arvstedt)
10d6b04ac8 support enabling clightning and lnd simultaneously (Erik Arvstedt)
ad7a519284 bitcoind: wait until RPC port is open (Erik Arvstedt)
5536b64fb3 lnd: wait until wallet is created (Erik Arvstedt)
6f2a55d63c lnd: wait until RPC port is open (Erik Arvstedt)
1868bef462 lnd: add option 'rpcPort' (Erik Arvstedt)
120e3e8cfe lnd postStart: suppress curl response output (Erik Arvstedt)
3e86637327 lnd postStart: poll for REST service availability (Erik Arvstedt)
795c51dc01 lnd postStart: make more idiomatic (Erik Arvstedt)
6e58beae8a lnd: use postStart option for script (Erik Arvstedt)
86167c6e6d clightning: wait until the RPC socket appears (Erik Arvstedt)
60c732a6a1 onion-chef: set RemainAfterExit, fix tor dependency (Erik Arvstedt)
2b9b3ba1c5 systemPackages: improve readability with shorter service references (Erik Arvstedt)
14ecb5511a liquid: add cli option (Erik Arvstedt)
cd5ed39b9c lnd: add cli option (Erik Arvstedt)
1833b15888 clightning: add cli option (Erik Arvstedt)
b90bf6691b add generate-secrets.service (Erik Arvstedt)
6447694214 add generate-secrets pkg (Erik Arvstedt)
e34093a8ac generate_secrets.sh: add opensslConf option (Erik Arvstedt)
9d14d5ba64 generate_secrets.sh: write secrets to working directory (Erik Arvstedt)
51fb054001 generate_secrets.sh: extract makepw command (Erik Arvstedt)
e3b47ce18a add setup-secrets.service (Erik Arvstedt)
437b268433 extract make-secrets.nix (Erik Arvstedt)
f9c29b9318 simplify secret definitions (Erik Arvstedt)
cd0fd6926b don't copy secret files to store during nixops deployment (Erik Arvstedt)
f0a36fe0c7 add 'nix-bitcoin-services' option (Erik Arvstedt)
7aaf30501c nix-bitcoin-services: simplify formatting (Erik Arvstedt)
760da232e0 add nix-bitcoin pkgs namespace (Erik Arvstedt)
6def181dbc add modules.nix (Erik Arvstedt)
3b842e5fe7 add nix-bitcoin-secrets.target (Erik Arvstedt)
bbf2bbc04a network.nix: simplify import of main config (Erik Arvstedt)
7e021a2629 simplify overlay.nix (Erik Arvstedt)
07dc3e04ac move bitcoinrpc group definition to bitcoind (Erik Arvstedt)
d61b185c3a simplify user and group definitions (Erik Arvstedt)
Pull request description:
The nix-bitcoin modules consist of three fundamental components:
1. a set of bitcoin-related modules for general use.
2. an opinionated configuration of these modules (`nix-bitcoin.nix`), to be deployed on a
dedicated machine.
3. machinery for nixops deployment.
This PR removes dependencies that reach from top to bottom in the list.
This means that 1. is now usable on its own and that 2. can be used without 3.
Besides improving nix-bitcoin's general usefulness, this
- simplifies testing. This PR includes a Travis-enabled modules test using the NixOS testing framework.
- paves the way for krops deployment.
- unlocks direct deployment in NixOS containers which allows for super fast experimentation.
### Details
Here are the unnecessary inter-component dependencies and how they're resolved by the commits. I'm using the numbering from the list above.
- `1. -> 3.` The modules (1.) use the nixops-specific (3.) `keys` group.
Resolved by `add nix-bitcoin-secrets.target`.
- `1. -> 3.` 1. requires nixops-specific key services.
Resolved by `add nix-bitcoin-secrets.target`.
- `1. -> 2.` bitcoind needs the bitcoinrpc group which is defined in `nix-bitcoin.nix` (2.).
Resolved by `move bitcoinrpc group definition to bitcoind`.
Further obstacles for standalone usage of 1.:
- We can't easily import 1. as a standalone module set.
Resolved by `add modules.nix`.
- Users of 1. shouldn't be forced to import nix-bitcoin's packages as top-level items in the pkgs namespace.
Resolved by `add nix-bitcoin pkgs namespace`.
### Non-nixops deployments
Commit `add setup-secrets.service` simplifies non-nixops deployment methods like containers, NixOS VMs or krops.
Secrets can now deployed as follows:
1. create local secrets.
2. transfer secrets to machine.
3. on the machine, `setup-secrets.service` creates extra secrets from `secrets.nix` and sets owner and
permissions for all secrets.
As krops integrates step 2. we now have all ingredients for automatic krops deployment.
The service is complicated by the creation of secrets like `bitcoin-rpcpassword` that are composed of attrs from `secrets.nix` instead of being simply backed by a file like `lnd_key`. We could simplify this by creating all secret files locally.
Running nix-bitcoin in NixOS containers gives you faster rebuild cycles when developing. [Here's](https://gist.github.com/5db4fa7dd3f1137920b58e39647116f6) an example.
### Test
The last commits starting with `clightning: add cli option` are testing-related and mostly fix non-critical bugs that were exposed by the test.
All `STABLE=1` builds from the Travis build matrix are implicit in the modules test.
Should we remove these individual builds?
Regarding commit `travis: cache all build outputs with cachix`:
To replace my cache with a cache that's owned by you (maybe named `nix-bitcoin-ci`), run
```
nix-shell -p travis --run 'travis encrypt CACHIX_SIGNING_KEY=... -r fort-nix/nix-bitcoin'
```
where `...` is the value of `secretKey` in `~/.config/cachix/cachix.dhall`. Let me know the travis secret and I'll fixup the commit.
### Docs
If you like the proposed changes, I'll add another PR with updates to the docs regarding the project layout, non-nixops deployment, and how to use nix-bitcoin within a larger NixOS config.
ACKs for top commit:
jonasnick:
ACK 187ff884db
Tree-SHA512: f4be65215c592a4f41bb7fa991a6d8d7c463cf631b88bf53051ca57ba280e7a60b8b09d0d1521345d5b656f844daa2166fff5d00a3105077c9e263465eacfb0a
Each secret file to be deployed is now backed by one local file.
This simplifies 'setup-secrets' and the secret definitions.
Also, with the old format it was not possible to add new secrets
to secrets.nix in a simple way.
Old secrets are automatically converted to the new format when running
nix-shell.
Using the new option 'nix-bitcoin.secrets', secrets are now directly
defined by the services that use them.
- set -e is implicit
- coreutils are in PATH and don't have to be explicitly referenced (echo is a shell builtin anyways)
- exit 0 is unneeded ('if' statements never fail)
This better fits the semantics of this unit and allows for easier
automated testing whether the service is active.
wantedBy = bindsTo = after = tor.service is the simplest way to ensure
that this unit is always running/restarted in lockstep with tor.
Previously, onion-chef would have stayed inactive in the case
that tor was stopped and then later restarted.
An executable is more robust to use than shell aliases.
This is also a preparation for commit 'add module test' because the
NixOS testing framework makes interactive aliases hard to use: It
unsets 'PS1' which is used by programs/bash/bash.nix to detect
interactive shells.
