GitHub ↗

Architecture

Two processes. The one exposed to the network has no way to touch the firewall.

Browser talks HTTPS to easywall-web, which runs unprivileged; easywall-web talks typed JSON over a Unix socket to easywall-core, which runs as root and speaks netlink to the nftables table inet easywall. Browser talks HTTPS to easywall-web, which runs unprivileged; easywall-web talks typed JSON over a Unix socket to easywall-core, which runs as root and speaks netlink to the nftables table inet easywall.

Who may do what

  easywall-web easywall-core
Runs as unprivileged user root, in the easywall group, with the capability set cut down to CAP_NET_ADMIN
Listens on HTTPS, port 12227 Unix socket only
Can reach nftables no yes, via netlink
Can run a shell no nft only: --check --file - to validate a custom rule, -f - to apply one
Holds sessions, templates rule state, acceptance timer

A flaw in form parsing or template rendering cannot reach the kernel, because the process holding that flaw has no kernel access to misuse. Reaching the firewall additionally requires a JSON command the typed protocol accepts and the core’s validation permits.

Why it is built this way. The original easywall — Python, Flask, iptables via subprocess — ran everything as one root process and passed user strings as command-line arguments. A web-layer bug was a firewall bug, and arguments could be injected. It was archived in 2022 after a CVE. Both root causes are gone here: the privileges are in a different process, and every typed rule reaches the kernel as a Go struct rather than a command line. Custom rules are the one string that still meets nft, and what keeps that safe is written out under Security.

Three rule sets

Editing writes to Staged. Applying copies Current to Backup and promotes Staged to Current. If the acceptance window expires, Backup is restored to Current. Editing writes to Staged. Applying copies Current to Backup and promotes Staged to Current. If the acceptance window expires, Backup is restored to Current.

Editing never touches the kernel. Saving writes to Staged; only Apply promotes it to Current, and Backup is what comes back if you do not confirm.

Applying, and the way back

State machine: editing leads to Staged, applying leads to Live, confirming within the window leads to Confirmed, and letting the window expire leads to Rolled back, from where the staged edits are still available. State machine: editing leads to Staged, applying leads to Live, confirming within the window leads to Confirmed, and letting the window expire leads to Rolled back, from where the staged edits are still available.

Applying is reversible by doing nothing. Not confirming — whether by choice or because you can no longer reach the page — restores the previous set.

   
Window 120 seconds by default, 10 to 3600
Where the timer lives the core daemon. Closing the tab or restarting easywall-web confirms nothing
Stopping the core mid-window counts as not confirming: the rules roll back before it exits
The same idea as commit confirmed on a Cisco router, or the at now + 5 minutes; iptables -F experienced operators type before a risky change — here it is the default rather than a habit

Step by step: Applying rules.

The socket protocol

Twenty-five message kinds, declared as Go structs on both sides. Adding an operation means adding a constant to both ends.

One exception, worth knowing: SaveRulesPayload.Rules is an interface{} that the core re-encodes and decodes into the type named by rule_type. An unknown rule_type is rejected. The decoded rules are validated before they are stored. The field itself is not typed at the protocol level — this page once claimed the whole protocol was.

Command Purpose
GET_RULES · SAVE_RULES read all three sets · write Staged
APPLY_RULES · ACCEPT promote Staged and start the timer · confirm
CANCEL_ACCEPTANCE end the open window now, without waiting out the timeout
GET_OPTIONS · SAVE_OPTIONS protection modules
GET_SETTINGS · SAVE_SETTINGS IPv6 and Docker
GET_SYSTEM · SAVE_SYSTEM acceptance window
GET_STATUS · GET_LOG dashboard · last 200 audit entries
GET_PACKET_LOG what the firewall refused — decoded packets from easywall’s NFLOG group, filtered, newest first
EXPORT_RULES · IMPORT_RULES backup and restore as JSON
VALIDATE_CUSTOM nft --check for the live editor
GET_APPLIED_CONFIG the options and network settings that went into the kernel with the rules that are in it
GET_USAGE what each port rule has carried, keyed by rule id, and when the figures were read
GET_HEALTH whether the firewall is doing what it says — three facts evaluated in order, plus the last self-test’s identity
PANIC · RESUME tear the table down and record it · end that and restore
LOG_EVENT one of thirteen login events, from a fixed enum, for the audit log
UPDATE_FEED a new version of one feed, fetched by the web process; the core re-checks every entry before it stores or loads any
GET_FEEDS per feed: entry counts, when it changed and was checked, packets dropped — never the entries

A message is at most 8 MiB each way. A longer request is answered request too large.

Full list: internal/shared/protocol.go.

What this does and does not protect you from

Defends against How
Web-layer vulnerabilities the web process has no kernel access
Argument injection into nft the apply path builds structs, not a command line
CSRF Origin and Sec-Fetch-Site checked on every unsafe method
Locking yourself out the acceptance window reverts on its own
Does not defend against Why
A compromised root account root owns the core
A kernel nftables vulnerability below easywall entirely
A legitimate admin making a bad rule the audit log records it, nothing prevents it

Next: Applying rules · Configuration · Security · How rules are ordered