mirror of
https://github.com/trailofbits/algo.git
synced 2025-08-19 11:13:05 +02:00
f668af22d0
* Fix VPN routing by adding output interface to NAT rules
The NAT rules were missing the output interface specification (-o eth0),
which caused routing failures on multi-homed systems (servers with multiple
network interfaces). Without specifying the output interface, packets might
not be NAT'd correctly.
Changes:
- Added -o {{ ansible_default_ipv4['interface'] }} to all NAT rules
- Updated both IPv4 and IPv6 templates
- Updated tests to verify output interface is present
- Added ansible_default_ipv4/ipv6 to test fixtures
This fixes the issue where VPN clients could connect but not route traffic
to the internet on servers with multiple network interfaces (like DigitalOcean
droplets with private networking enabled).
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix VPN routing by adding output interface to NAT rules
On multi-homed systems (servers with multiple network interfaces or multiple IPs
on one interface), MASQUERADE rules need to specify which interface to use for
NAT. Without the output interface specification, packets may not be routed correctly.
This fix adds the output interface to all NAT rules:
-A POSTROUTING -s [vpn_subnet] -o eth0 -j MASQUERADE
Changes:
- Modified roles/common/templates/rules.v4.j2 to include output interface
- Modified roles/common/templates/rules.v6.j2 for IPv6 support
- Added tests to verify output interface is present in NAT rules
- Added ansible_default_ipv4/ipv6 variables to test fixtures
For deployments on providers like DigitalOcean where MASQUERADE still fails
due to multiple IPs on the same interface, users can enable the existing
alternative_ingress_ip option in config.cfg to use explicit SNAT.
Testing:
- Verified on live servers
- All unit tests pass (67/67)
- Mutation testing confirms test coverage
This fixes VPN connectivity on servers with multiple interfaces while
remaining backward compatible with single-interface deployments.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix dnscrypt-proxy not listening on VPN service IPs
Problem: dnscrypt-proxy on Ubuntu uses systemd socket activation by default,
which overrides the configured listen_addresses in dnscrypt-proxy.toml.
The socket only listens on 127.0.2.1:53, preventing VPN clients from
resolving DNS queries through the configured service IPs.
Solution: Disable and mask the dnscrypt-proxy.socket unit to allow
dnscrypt-proxy to bind directly to the VPN service IPs specified in
its configuration file.
This fixes DNS resolution for VPN clients on Ubuntu 20.04+ systems.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Apply Python linting and formatting
- Run ruff check --fix to fix linting issues
- Run ruff format to ensure consistent formatting
- All tests still pass after formatting changes
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Restrict DNS access to VPN clients only
Security fix: The firewall rule for DNS was accepting traffic from any
source (0.0.0.0/0) to the local DNS resolver. While the service IP is
on the loopback interface (which normally isn't routable externally),
this could be a security risk if misconfigured.
Changed firewall rules to only accept DNS traffic from VPN subnets:
- INPUT rule now includes -s {{ subnets }} to restrict source IPs
- Applied to both IPv4 and IPv6 rules
- Added test to verify DNS is properly restricted
This ensures the DNS resolver is only accessible to connected VPN
clients, not the entire internet.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix dnscrypt-proxy service startup with masked socket
Problem: dnscrypt-proxy.service has a dependency on dnscrypt-proxy.socket
through the TriggeredBy directive. When we mask the socket before starting
the service, systemd fails with "Unit dnscrypt-proxy.socket is masked."
Solution:
1. Override the service to remove socket dependency (TriggeredBy=)
2. Reload systemd daemon immediately after override changes
3. Start the service (which now doesn't require the socket)
4. Only then disable and mask the socket
This ensures dnscrypt-proxy can bind directly to the configured IPs
without socket activation, while preventing the socket from being
re-enabled by package updates.
Changes:
- Added TriggeredBy= override to remove socket dependency
- Added explicit daemon reload after service overrides
- Moved socket masking to after service start in main.yml
- Fixed YAML formatting issues
Testing: Deployment now succeeds with dnscrypt-proxy binding to VPN IPs
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix dnscrypt-proxy by not masking the socket
Problem: Masking dnscrypt-proxy.socket prevents the service from starting
because the service has Requires=dnscrypt-proxy.socket dependency.
Solution: Simply stop and disable the socket without masking it. This
prevents socket activation while allowing the service to start and bind
directly to the configured IPs.
Changes:
- Removed socket masking (just disable it)
- Moved socket disabling before service start
- Removed invalid systemd directives from override
Testing: Confirmed dnscrypt-proxy now listens on VPN service IPs
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Use systemd socket activation properly for dnscrypt-proxy
Instead of fighting systemd socket activation, configure it to listen
on the correct VPN service IPs. This is more systemd-native and reliable.
Changes:
- Create socket override to listen on VPN IPs instead of localhost
- Clear default listeners and add VPN service IPs
- Use empty listen_addresses in dnscrypt-proxy.toml for socket activation
- Keep socket enabled and let systemd manage the activation
- Add handler for restarting socket when config changes
Benefits:
- Works WITH systemd instead of against it
- Survives package updates better
- No dependency conflicts
- More reliable service management
This approach is cleaner than disabling socket activation entirely and
ensures dnscrypt-proxy is accessible to VPN clients on the correct IPs.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Document debugging lessons learned in CLAUDE.md
Added comprehensive debugging guidance based on our troubleshooting session:
- VPN connectivity troubleshooting order (DNS first!)
- systemd socket activation best practices
- Common deployment failures and solutions
- Time wasters to avoid (lessons learned the hard way)
- Multi-homed system considerations
- Testing notes for DigitalOcean
These additions will help future debugging sessions avoid the same
rabbit holes and focus on the most likely issues first.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix DNS resolution for VPN clients by enabling route_localnet
The issue was that dnscrypt-proxy listens on a special loopback IP
(randomly generated in 172.16.0.0/12 range) which wasn't accessible
from VPN clients. This fix:
1. Enables net.ipv4.conf.all.route_localnet sysctl to allow routing
to loopback IPs from other interfaces
2. Ensures dnscrypt-proxy socket is properly restarted when its
configuration changes
3. Adds proper handler flushing after socket configuration updates
This allows VPN clients to reach the DNS resolver at the local_service_ip
address configured on the loopback interface.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Improve security by using interface-specific route_localnet
Instead of enabling route_localnet globally (net.ipv4.conf.all.route_localnet),
this change enables it only on the specific interfaces that need it:
- WireGuard interface (wg0) for WireGuard VPN clients
- Main network interface (eth0/etc) for IPsec VPN clients
This minimizes the security impact by restricting loopback routing to only
the VPN interfaces, preventing other interfaces from being able to route
to loopback addresses.
The interface-specific approach provides the same functionality (allowing
VPN clients to reach the DNS resolver on the local_service_ip) while
reducing the potential attack surface.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Revert to global route_localnet to fix deployment failure
The interface-specific route_localnet approach failed because:
- WireGuard interface (wg0) doesn't exist until the service starts
- We were trying to set the sysctl before the interface was created
- This caused deployment failures with "No such file or directory"
Reverting to the global setting (net.ipv4.conf.all.route_localnet=1) because:
- It always works regardless of interface creation timing
- VPN users are trusted (they have our credentials)
- Firewall rules still restrict access to only port 53
- The security benefit of interface-specific settings is minimal
- The added complexity isn't worth the marginal security improvement
This ensures reliable deployments while maintaining the DNS resolution fix.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Fix dnscrypt-proxy socket restart and remove problematic BPF hardening
Two important fixes:
1. Fix dnscrypt-proxy socket not restarting with new configuration
- The socket wasn't properly restarting when its override config changed
- This caused DNS to listen on wrong IP (127.0.2.1 instead of local_service_ip)
- Now directly restart the socket when configuration changes
- Add explicit daemon reload before restarting
2. Remove BPF JIT hardening that causes deployment errors
- The net.core.bpf_jit_enable sysctl isn't available on all kernels
- It was causing "Invalid argument" errors during deployment
- This was optional security hardening with minimal benefit
- Removing it eliminates deployment errors for most users
These fixes ensure reliable DNS resolution for VPN clients and clean
deployments without error messages.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
* Update CLAUDE.md with comprehensive debugging lessons learned
Based on our extensive debugging session, this update adds critical documentation:
## DNS Architecture and Troubleshooting
- Explained the local_service_ip design and why it requires route_localnet
- Added detailed DNS debugging methodology with exact steps in order
- Documented systemd socket activation complexities and common mistakes
- Added specific commands to verify DNS is working correctly
## Architectural Decisions
- Added new section explaining trade-offs in Algo's design choices
- Documented why local_service_ip uses loopback instead of alternatives
- Explained iptables-legacy vs iptables-nft backend choice
## Enhanced Debugging Guidance
- Expanded troubleshooting with exact commands and expected outputs
- Added warnings about configuration changes that need restarts
- Documented socket activation override requirements in detail
- Added common pitfalls like interface-specific sysctls
## Time Wasters Section
- Added new lessons learned from this debugging session
- Interface-specific route_localnet (fails before interface exists)
- DNAT for loopback addresses (doesn't work)
- BPF JIT hardening (causes errors on many kernels)
This documentation will help future maintainers avoid the same debugging
rabbit holes and understand why things are designed the way they are.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
|
||
|---|---|---|
| .. | ||
| annotate-test-failure.sh | ||
| lint.sh | ||
| test-templates.sh | ||
| track-test-effectiveness.py | ||