3x-ui/internal/amneziawg/manager.go

package amneziawg

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"fmt"
	"maps"
	"net"
	"net/netip"
	"os"
	"os/exec"
	"path/filepath"
	"slices"
	"strconv"
	"strings"
	"sync"
	"time"

	"github.com/mhsanaei/3x-ui/v3/internal/database/model"
	"github.com/mhsanaei/3x-ui/v3/internal/logger"
)

// configDir is where awg-quick expects to find <interface>.conf, matching
// the AmneziaWG DKMS package's own layout.
const configDir = "/etc/amnezia/amneziawg"

// onlineWindow is how recent a peer's last handshake must be to count it as
// online, matching the typical WireGuard rekey interval (every 120s) plus
// margin.
const onlineWindow = 180 * time.Second

// InstanceFromInbound derives a desired Instance from an AmneziaWG inbound,
// building one peer per active client. Returns false when the inbound is not
// a usable AmneziaWG inbound (wrong protocol, unparseable settings, or no
// server block) or has no enabled peer to serve — mirroring
// mtproto.InstanceFromInbound, which skips the sidecar entirely rather than
// run it with nothing to serve.
func InstanceFromInbound(ib *model.Inbound) (Instance, bool) {
	if ib == nil || ib.Protocol != model.AmneziaWG {
		return Instance{}, false
	}
	var parsed InboundSettings
	if err := json.Unmarshal([]byte(ib.Settings), &parsed); err != nil || parsed.Server == nil {
		return Instance{}, false
	}
	server := parsed.Server

	peers := make([]Peer, 0, len(parsed.Clients))
	for _, c := range parsed.Clients {
		if !c.Enable || c.PublicKey == "" || len(c.AllowedIPs) == 0 {
			continue
		}
		peers = append(peers, Peer{
			Email:          c.Email,
			PublicKey:      c.PublicKey,
			PresharedKey:   c.PreSharedKey,
			AllowedIPs:     c.AllowedIPs,
			ForwardedPorts: c.ForwardedPorts,
		})
	}
	if len(peers) == 0 {
		return Instance{}, false
	}

	addresses := []string{serverAddress(server.SubnetIP, server.SubnetCIDR)}
	if server.IPv6Enabled {
		if v6, ok := serverAddressV6(server.IPv6Subnet); ok {
			addresses = append(addresses, v6)
		}
	}

	return Instance{
		Id:                    ib.Id,
		Tag:                   ib.Tag,
		InterfaceName:         interfaceNameForID(ib.Id),
		ListenPort:            ib.Port,
		PrivateKey:            server.PrivateKey,
		PublicKey:             server.PublicKey,
		Address:               addresses,
		MTU:                   server.MTU,
		Obfuscation:           server.Obfuscation(),
		Peers:                 peers,
		ExternalInterface:     server.ExternalInterface,
		IPv6Enabled:           server.IPv6Enabled,
		IPv6ExternalInterface: server.IPv6ExternalInterface,
		RouteThroughXray:      server.RouteThroughXray,
	}, true
}

// interfaceNameForID derives the OS-level interface name for an inbound, e.g.
// "awg42".
func interfaceNameForID(id int) string {
	return fmt.Sprintf("awg%d", id)
}

// serverAddress returns the server's own tunnel address for a subnet base,
// e.g. "10.8.1.1/24" for base "10.8.1.0" or "10.8.1.5". The server always
// holds the first usable host of the network subnetIP/cidr actually
// describes -- derived via netip rather than assuming subnetIP already ends
// in ".0", so a subnetIP that isn't a bare network address (a typo, or a
// manually edited value) can never collide with peer addresses, which are
// allocated starting from the network's second host upward (see
// allocateWireguardAddress). Falls back to the previous literal behavior
// only if subnetIP/cidr doesn't parse as an IPv4 network at all -- normal
// saves never reach that path since ValidateSubnetIPv4 already rejects it.
func serverAddress(subnetIP string, cidr int) string {
	if cidr <= 0 {
		cidr = 24
	}
	// A /32 has no host bits at all -- "first usable host" is meaningless,
	// and Next() would step outside the block entirely -- so a single-host
	// base is used exactly as given, same as before this fix.
	prefix, err := netip.ParsePrefix(fmt.Sprintf("%s/%d", subnetIP, cidr))
	if err != nil || !prefix.Addr().Is4() || cidr >= 32 {
		return fmt.Sprintf("%s/%d", subnetIP, cidr)
	}
	host := prefix.Masked().Addr().Next()
	return fmt.Sprintf("%s/%d", host, cidr)
}

// serverAddressV6 returns the server's own IPv6 tunnel address for a subnet
// CIDR (e.g. "fd86:ea04:1115::1/64" for "fd86:ea04:1115::/64"), the first
// usable host in the prefix. ok is false when subnetCIDR is empty or not a
// valid IPv6 prefix.
func serverAddressV6(subnetCIDR string) (addr string, ok bool) {
	prefix, err := netip.ParsePrefix(subnetCIDR)
	if err != nil || !prefix.Addr().Is6() {
		return "", false
	}
	host := prefix.Masked().Addr().Next()
	return fmt.Sprintf("%s/%d", host, prefix.Bits()), true
}

// structuralFingerprint changes whenever a value that requires a full
// interface bounce (awg-quick down + up) changes.
func (inst Instance) structuralFingerprint() string {
	o := inst.Obfuscation
	parts := []string{
		inst.InterfaceName,
		strconv.Itoa(inst.ListenPort),
		inst.PrivateKey,
		strings.Join(inst.Address, ","),
		strconv.Itoa(inst.MTU),
		strconv.Itoa(o.Jc), strconv.Itoa(o.Jmin), strconv.Itoa(o.Jmax),
		strconv.Itoa(o.S1), strconv.Itoa(o.S2), strconv.Itoa(o.S3), strconv.Itoa(o.S4),
		o.H1, o.H2, o.H3, o.H4, o.I1,
		inst.ExternalInterface,
		strconv.FormatBool(inst.IPv6Enabled),
		inst.IPv6ExternalInterface,
		strconv.FormatBool(inst.RouteThroughXray),
	}
	return strings.Join(parts, "|")
}

// peersFingerprint identifies the reloadable peer set regardless of order, so
// a reordered clients array in the stored settings does not read as a
// change. It moves whenever a peer is added, removed, disabled, re-keyed, or
// re-addressed — all of which `awg syncconf` applies in place. Deliberately
// excludes ForwardedPorts: those live in PostUp/PostDown, not the WireGuard
// peer table, so a ports-only change needs hostRulesFingerprint's full
// bounce instead of a syncconf reload.
func (inst Instance) peersFingerprint() string {
	pairs := make([]string, 0, len(inst.Peers))
	for _, p := range inst.Peers {
		pairs = append(pairs, fmt.Sprintf("%s=%s;psk=%s;ips=%s", p.Email, p.PublicKey, p.PresharedKey, strings.Join(p.AllowedIPs, ",")))
	}
	slices.Sort(pairs)
	return strings.Join(pairs, "|")
}

// hostRulesFingerprint identifies per-peer state that only ever takes effect
// through PostUp/PostDown shell rules — forwarded ports (whose DNAT rules
// are keyed on the peer's IPv4 address, the same as the TPROXY rule below);
// when RouteThroughXray is on, every peer's IPv4 address (the TPROXY rule
// into this instance's own Xray bridge is keyed on it); and when IPv6 is
// enabled, the peer's IPv6 address (its NDP-proxy PostUp/PostDown entry) —
// rather than the WireGuard peer table itself. It is checked separately from
// peersFingerprint because `awg syncconf` never re-runs PostUp/PostDown, so
// a change here must force a full interface bounce (ensureRestart) to
// actually take effect, unlike a key-only change that syncconf can apply in
// place. The IPv4 component is included whenever RouteThroughXray is on OR
// the peer has forwarded ports — either one means PostUp/PostDown text is
// keyed on that address, so a re-IP with either feature off must still force
// a bounce (otherwise the old DNAT/TPROXY rule survives pointed at an
// address the reconciler is now free to hand to a different peer). The IPv6
// component stays IPv6Enabled-gated only, matching the single feature that
// reads it. Skipping both entirely when neither applies preserves the
// syncconf fast path for a plain instance's peer add/remove/re-IP.
func (inst Instance) hostRulesFingerprint() string {
	pairs := make([]string, 0, len(inst.Peers))
	for _, p := range inst.Peers {
		v := fmt.Sprintf("%s=fwd:%s", p.Email, p.ForwardedPorts)
		if inst.RouteThroughXray || p.ForwardedPorts != "" {
			v += ";ip:" + FirstIPv4(p.AllowedIPs)
		}
		if inst.IPv6Enabled {
			v += ";ip6:" + firstIPv6(p.AllowedIPs)
		}
		pairs = append(pairs, v)
	}
	slices.Sort(pairs)
	return strings.Join(pairs, "|")
}

// peerCounters is the last-seen cumulative transfer counters for one peer,
// used to compute per-poll deltas the same way mtproto tracks per-secret
// counters.
type peerCounters struct {
	rx int64
	tx int64
}

type managed struct {
	inst         Instance
	structuralFP string
	peersFP      string
	hostRulesFP  string
	last         map[string]peerCounters // keyed by peer public key
}

// Manager owns the set of running AmneziaWG interfaces keyed by inbound id.
type Manager struct {
	mu     sync.Mutex
	ifaces map[int]*managed
	// swept records that the one-time startup cleanup of orphaned interfaces
	// (survivors of a previous x-ui run) has already run.
	swept bool
}

var (
	managerOnce sync.Once
	manager     *Manager
)

// GetManager returns the process-wide AmneziaWG manager singleton.
func GetManager() *Manager {
	managerOnce.Do(func() {
		manager = &Manager{ifaces: map[int]*managed{}}
	})
	return manager
}

// ensureAction is what ensureLocked must do to move a running interface to a
// desired instance: leave it alone, hot-reload just its peers, or fully
// bounce it.
type ensureAction int

const (
	ensureNoop ensureAction = iota
	ensureReload
	ensureRestart
)

// ensureActionFor decides how to apply a desired instance to the currently
// managed interface. A structural change, a host-rules change (forwarded
// ports, or simply a peer's presence/IP — its always-on TPROXY rule only
// lives in PostUp/PostDown), or a down interface all force a restart; a
// peers-only change (keys only, no IP/presence change) is a candidate for
// an in-place `syncconf`; identical fingerprints on an up interface need
// nothing.
func ensureActionFor(up bool, curStructFP, curHostRulesFP, curPeersFP, newStructFP, newHostRulesFP, newPeersFP string) ensureAction {
	if !up || curStructFP != newStructFP || curHostRulesFP != newHostRulesFP {
		return ensureRestart
	}
	if curPeersFP != newPeersFP {
		return ensureReload
	}
	return ensureNoop
}

// Ensure brings one interface to its desired state, or restarts/reloads it
// when its configuration changed. A no-op when it already matches.
func (m *Manager) Ensure(inst Instance) error {
	m.mu.Lock()
	defer m.mu.Unlock()
	return m.ensureLocked(inst)
}

func (m *Manager) ensureLocked(inst Instance) error {
	structFP := inst.structuralFingerprint()
	hostRulesFP := inst.hostRulesFingerprint()
	peersFP := inst.peersFingerprint()

	cur, exists := m.ifaces[inst.Id]
	action := ensureRestart
	if exists {
		action = ensureActionFor(isInterfaceUp(cur.inst.InterfaceName), cur.structuralFP, cur.hostRulesFP, cur.peersFP, structFP, hostRulesFP, peersFP)
	}

	switch action {
	case ensureNoop:
		cur.inst = inst
		return nil
	case ensureReload:
		if err := writeConfigFile(inst); err != nil {
			return err
		}
		if err := syncConfig(inst); err != nil {
			return err
		}
	case ensureRestart:
		// Checked against the interface's actual kernel state, not `exists`:
		// after an ungraceful exit (kill -9, OOM, panic) the previous
		// process's interface can still be up even though this fresh
		// Manager has never seen it (exists is always false on a cold
		// start). Skipping the teardown in that case would send
		// interfaceUp straight into "ip link add" against a name that
		// already exists, which fails and leaves this inbound stuck
		// retrying every reconcile forever.
		if isInterfaceUp(inst.InterfaceName) {
			_ = interfaceDown(inst.InterfaceName)
		}
		if err := writeConfigFile(inst); err != nil {
			return err
		}
		if err := interfaceUp(inst.InterfaceName); err != nil {
			return err
		}
		logger.Infof("amneziawg: started interface %s for inbound %d", inst.InterfaceName, inst.Id)
	}

	last := map[string]peerCounters{}
	if exists {
		last = nextTrafficBaseline(action, cur.last)
	}
	m.ifaces[inst.Id] = &managed{inst: inst, structuralFP: structFP, hostRulesFP: hostRulesFP, peersFP: peersFP, last: last}
	return nil
}

// nextTrafficBaseline decides what per-peer traffic counters ensureLocked
// should carry into the next managed entry. Only a reload (awg syncconf)
// preserves the kernel's own per-peer transfer counters; a full down+up
// zeroes them. Carrying the old baseline forward after a restart would make
// the next CollectTraffic compute a large negative delta (clamped to 0 by
// the caller), silently discarding whatever the peers transferred since the
// previous poll instead of just resuming the count from zero.
func nextTrafficBaseline(action ensureAction, prev map[string]peerCounters) map[string]peerCounters {
	if action == ensureReload {
		return prev
	}
	return map[string]peerCounters{}
}

// Remove tears down and forgets the interface for an inbound id.
func (m *Manager) Remove(id int) {
	m.mu.Lock()
	defer m.mu.Unlock()
	if cur, ok := m.ifaces[id]; ok {
		_ = interfaceDown(cur.inst.InterfaceName)
		removeConfigFile(cur.inst.InterfaceName)
		delete(m.ifaces, id)
		logger.Infof("amneziawg: stopped interface %s for inbound %d", cur.inst.InterfaceName, id)
	}
}

// sweepOrphansLocked tears down any AmneziaWG interface and config file left
// behind by a previous x-ui process whose inbound is no longer in the
// current desired set — most commonly because it was deleted from the
// database entirely while the panel was down, so it will never again appear
// in any future Reconcile call and would otherwise never be discovered (it
// has no entry in m.ifaces for the per-id cleanup loop below to catch,
// because that map always starts empty on a fresh process). Runs once per
// process lifetime, mirroring mtproto.Manager.sweepOrphansLocked.
//
// Deliberately only called from Reconcile, not Ensure: Ensure only ever
// carries a single instance, and a `want` set of just that one id would
// misidentify every other still-desired-but-not-yet-reconciled-this-process
// interface as an orphan. A crashed-but-still-wanted interface is instead
// recovered normally by ensureLocked's ensureRestart branch, which checks
// the interface's actual kernel state rather than this manager's in-memory
// bookkeeping.
func (m *Manager) sweepOrphansLocked(want map[int]struct{}) {
	if m.swept {
		return
	}
	entries, err := os.ReadDir(configDir)
	if err != nil {
		// Left false on purpose: a transient error (the directory not existing
		// yet, a momentary filesystem hiccup) should let the next Reconcile
		// tick retry the sweep, rather than permanently disabling it for this
		// process's whole lifetime over a failure that may not recur.
		return
	}
	m.swept = true
	names := make([]string, 0, len(entries))
	for _, entry := range entries {
		if !entry.IsDir() {
			names = append(names, entry.Name())
		}
	}
	for _, ifaceName := range orphanedInterfaces(names, want) {
		if isInterfaceUp(ifaceName) {
			_ = interfaceDown(ifaceName)
			logger.Warningf("amneziawg: tore down orphaned interface %s (its inbound no longer exists)", ifaceName)
		}
		removeConfigFile(ifaceName)
	}
}

// orphanedInterfaces returns the interface names among confFileNames (the
// basenames of configDir's entries) whose parsed inbound id is not present
// in want — the pure decision sweepOrphansLocked acts on.
func orphanedInterfaces(confFileNames []string, want map[int]struct{}) []string {
	var out []string
	for _, name := range confFileNames {
		if !strings.HasSuffix(name, ".conf") {
			continue
		}
		ifaceName := strings.TrimSuffix(name, ".conf")
		id, ok := inboundIDForInterfaceName(ifaceName)
		if !ok {
			continue
		}
		if _, wanted := want[id]; wanted {
			continue
		}
		out = append(out, ifaceName)
	}
	return out
}

// inboundIDForInterfaceName parses the inbound id back out of an interface
// name produced by interfaceNameForID, e.g. "awg42" -> 42, ok=true. Requires
// the suffix to be all decimal digits so a stray or hand-crafted file name
// (e.g. "awg-1.conf") can never resolve to a negative id.
func inboundIDForInterfaceName(name string) (int, bool) {
	suffix, ok := strings.CutPrefix(name, "awg")
	if !ok || suffix == "" {
		return 0, false
	}
	for _, r := range suffix {
		if r < '0' || r > '9' {
			return 0, false
		}
	}
	id, err := strconv.Atoi(suffix)
	if err != nil {
		return 0, false
	}
	return id, true
}

// Reconcile drives the running set toward the desired instances: it tears
// down interfaces that are no longer wanted and ensures the rest. Used at
// boot and periodically to recover from crashes or an out-of-band `awg-quick
// down`.
func (m *Manager) Reconcile(desired []Instance) {
	m.mu.Lock()
	defer m.mu.Unlock()
	want := make(map[int]struct{}, len(desired))
	for _, inst := range desired {
		want[inst.Id] = struct{}{}
	}
	m.sweepOrphansLocked(want)
	for id, cur := range m.ifaces {
		if _, ok := want[id]; !ok {
			_ = interfaceDown(cur.inst.InterfaceName)
			removeConfigFile(cur.inst.InterfaceName)
			delete(m.ifaces, id)
			logger.Infof("amneziawg: stopped interface %s for removed inbound %d", cur.inst.InterfaceName, id)
		}
	}
	for _, inst := range desired {
		if err := m.ensureLocked(inst); err != nil {
			logger.Warningf("amneziawg: reconcile failed for inbound %d: %v", inst.Id, err)
		}
	}
}

// StopAll tears down every managed interface. Called on panel shutdown.
func (m *Manager) StopAll() {
	m.mu.Lock()
	defer m.mu.Unlock()
	for id, cur := range m.ifaces {
		_ = interfaceDown(cur.inst.InterfaceName)
		delete(m.ifaces, id)
	}
}

// HasRunning reports whether any managed interface is currently up.
func (m *Manager) HasRunning() bool {
	m.mu.Lock()
	defer m.mu.Unlock()
	for _, cur := range m.ifaces {
		if isInterfaceUp(cur.inst.InterfaceName) {
			return true
		}
	}
	return false
}

// Traffic is a per-peer traffic delta scraped from `awg show <iface> dump`.
// Tag is the owning inbound's tag and Email is the client the bytes belong
// to.
type Traffic struct {
	Tag   string
	Email string
	Up    int64
	Down  int64
}

// CollectTraffic polls `awg show <iface> dump` for every running interface
// and returns the per-peer byte deltas since the previous poll, plus the
// emails of peers with a handshake inside onlineWindow.
func (m *Manager) CollectTraffic() ([]Traffic, []string) {
	type snap struct {
		id   int
		inst Instance
		last map[string]peerCounters
		// entry is the exact *managed snapshotted below, kept so the
		// write-back can detect a concurrent ensureRestart/ensureReload
		// (which replaces the map entry with a fresh pointer, see
		// ensureLocked) that happened while getPeerStats ran lock-free.
		entry *managed
	}
	m.mu.Lock()
	snaps := make([]snap, 0, len(m.ifaces))
	for id, cur := range m.ifaces {
		lastCopy := make(map[string]peerCounters, len(cur.last))
		maps.Copy(lastCopy, cur.last)
		snaps = append(snaps, snap{id: id, inst: cur.inst, last: lastCopy, entry: cur})
	}
	m.mu.Unlock()

	var out []Traffic
	var online []string
	now := time.Now()

	for _, s := range snaps {
		stats, err := getPeerStats(s.inst.InterfaceName)
		if err != nil {
			continue
		}
		emailByKey := make(map[string]string, len(s.inst.Peers))
		for _, p := range s.inst.Peers {
			emailByKey[p.PublicKey] = p.Email
		}

		newLast := make(map[string]peerCounters, len(stats))
		for _, st := range stats {
			email, ok := emailByKey[st.publicKey]
			if !ok || email == "" {
				continue
			}
			newLast[st.publicKey] = peerCounters{rx: st.rx, tx: st.tx}
			if st.latestHandshake > 0 && now.Sub(time.Unix(st.latestHandshake, 0)) < onlineWindow {
				online = append(online, email)
			}
			prev, had := s.last[st.publicKey]
			if !had {
				continue
			}
			du := st.rx - prev.rx // client upload = bytes the server received
			dd := st.tx - prev.tx // client download = bytes the server sent
			if du < 0 {
				du = 0
			}
			if dd < 0 {
				dd = 0
			}
			if du > 0 || dd > 0 {
				out = append(out, Traffic{Tag: s.inst.Tag, Email: email, Up: du, Down: dd})
			}
		}

		m.mu.Lock()
		// Only write back if this is still the exact entry snapshotted above:
		// getPeerStats ran without the lock held, so ensureLocked could have
		// restarted (or reloaded) this same interface in the meantime,
		// replacing the map entry with a fresh *managed and, for a restart,
		// resetting last to empty (kernel counters zero on down+up). Writing
		// newLast back over that unconditionally would silently resurrect the
		// pre-restart counters as the new baseline, making the next poll
		// compute a negative delta and clamp a real poll's worth of traffic
		// to zero.
		if cur, ok := m.ifaces[s.id]; ok && cur == s.entry {
			cur.last = newLast
		}
		m.mu.Unlock()
	}
	return out, online
}

// --- config rendering ---

// generateServerConfig builds the awg-quick .conf content for an interface:
// its own [Interface] block (keys, address, obfuscation, NAT PostUp/PostDown)
// followed by one [Peer] block per client.
func generateServerConfig(inst Instance) string {
	var b strings.Builder

	b.WriteString("[Interface]\n")
	fmt.Fprintf(&b, "PrivateKey = %s\n", sanitizeConfigValue(inst.PrivateKey))
	if len(inst.Address) > 0 {
		fmt.Fprintf(&b, "Address = %s\n", strings.Join(inst.Address, ", "))
	}
	fmt.Fprintf(&b, "ListenPort = %d\n", inst.ListenPort)
	if inst.MTU > 0 {
		fmt.Fprintf(&b, "MTU = %d\n", inst.MTU)
	}
	writeObfuscation(&b, inst.Obfuscation)

	ext := inst.ExternalInterface
	if ext == "" {
		ext = detectDefaultInterface()
	}
	postUp, postDown := defaultPostUpDown(inst, ext)
	fmt.Fprintf(&b, "PostUp = %s\n", postUp)
	fmt.Fprintf(&b, "PostDown = %s\n", postDown)

	for _, p := range inst.Peers {
		b.WriteString("\n[Peer]\n")
		if p.Email != "" {
			fmt.Fprintf(&b, "# %s\n", sanitizeConfigValue(p.Email))
		}
		fmt.Fprintf(&b, "PublicKey = %s\n", sanitizeConfigValue(p.PublicKey))
		if p.PresharedKey != "" {
			fmt.Fprintf(&b, "PresharedKey = %s\n", sanitizeConfigValue(p.PresharedKey))
		}
		fmt.Fprintf(&b, "AllowedIPs = %s\n", strings.Join(p.AllowedIPs, ", "))
	}

	return b.String()
}

// sanitizeConfigValue strips newlines, carriage returns, and other control
// characters from a value about to be interpolated into the generated
// .conf. ValidateConfigValue rejects these at save time, but a row that
// predates that validation (an upgrade, a node sync, a restored backup, a
// direct DB edit) would otherwise still reach awg-quick's parser, where a
// newline lets a later line re-open a new section and smuggle in a hook
// awg-quick executes as root. This is the render-time backstop; it
// silently drops the offending bytes rather than failing the whole config
// build, matching how hOrDefault degrades a blank H value instead of
// emitting an invalid line.
func sanitizeConfigValue(v string) string {
	return strings.Map(func(r rune) rune {
		if r == '\n' || r == '\r' || r < 0x20 || r == 0x7f {
			return -1
		}
		return r
	}, v)
}

// writeObfuscation writes the AmneziaWG obfuscation parameters that must be
// identical on both ends of a tunnel. S3/S4 and I1 are emitted only when set,
// so a plain 1.x-equivalent set (S3=S4=0, I1="") produces the classic
// generator's output; a 2.0 set adds the extra padding, header ranges and CPS
// packet.
func writeObfuscation(b *strings.Builder, o Obfuscation20) {
	fmt.Fprintf(b, "Jc = %d\n", o.Jc)
	fmt.Fprintf(b, "Jmin = %d\n", o.Jmin)
	fmt.Fprintf(b, "Jmax = %d\n", o.Jmax)
	fmt.Fprintf(b, "S1 = %d\n", o.S1)
	fmt.Fprintf(b, "S2 = %d\n", o.S2)
	if o.S3 > 0 {
		fmt.Fprintf(b, "S3 = %d\n", o.S3)
	}
	if o.S4 > 0 {
		fmt.Fprintf(b, "S4 = %d\n", o.S4)
	}
	fmt.Fprintf(b, "H1 = %s\n", hOrDefault(o.H1, "1"))
	fmt.Fprintf(b, "H2 = %s\n", hOrDefault(o.H2, "2"))
	fmt.Fprintf(b, "H3 = %s\n", hOrDefault(o.H3, "3"))
	fmt.Fprintf(b, "H4 = %s\n", hOrDefault(o.H4, "4"))
	if o.I1 != "" {
		fmt.Fprintf(b, "I1 = %s\n", sanitizeConfigValue(o.I1))
	}
}

// hOrDefault returns def when v is blank, guarding against an empty H value
// (which would emit an invalid "H1 = " line) on legacy/partial records.
func hOrDefault(v, def string) string {
	if strings.TrimSpace(v) == "" {
		return def
	}
	return v
}

// defaultPostUpDown returns NAT + forwarding rules: MASQUERADE the tunnel
// subnet out the external interface, accept forwarded traffic in both
// directions, and — when the instance has IPv6 enabled — the IPv6-forward
// rules, proxy_ndp sysctl, and one `ip -6 neigh add proxy` entry per enabled
// peer with an IPv6 address, so upstream routers see each client's IPv6 as
// directly reachable on the LAN without NAT66. Also emits DNAT+FORWARD rules
// for each enabled peer with a non-empty ForwardedPorts spec, and — only
// when the instance has RouteThroughXray enabled — a mangle-table TPROXY
// rule redirecting every peer's traffic into this instance's own Xray
// bridge (see EgressPortForInbound), plus the one-time policy route TPROXY
// needs to deliver it there. RouteThroughXray is off by default: a plain
// AmneziaWG tunnel has no Xray dependency at all unless the admin opts in.
// When it is on, it is entirely up to the admin's own Xray Routing rules
// (targeting this inbound's own tag, which injectAmneziawgEgress reuses for
// the bridge) whether that traffic ever actually goes anywhere beyond
// Xray's default routing.
func defaultPostUpDown(inst Instance, ext string) (postUp, postDown string) {
	iface := inst.InterfaceName
	up := []string{
		fmt.Sprintf("iptables -A FORWARD -i %s -j ACCEPT", iface),
		fmt.Sprintf("iptables -A FORWARD -o %s -j ACCEPT", iface),
	}
	down := []string{
		fmt.Sprintf("iptables -D FORWARD -i %s -j ACCEPT", iface),
		fmt.Sprintf("iptables -D FORWARD -o %s -j ACCEPT", iface),
	}
	if subnet := firstAddress(inst.Address); subnet != "" && ext != "" {
		up = append([]string{fmt.Sprintf("iptables -t nat -A POSTROUTING -s %s -o %s -j MASQUERADE", subnet, ext)}, up...)
		down = append([]string{fmt.Sprintf("iptables -t nat -D POSTROUTING -s %s -o %s -j MASQUERADE", subnet, ext)}, down...)
	}

	if inst.IPv6Enabled {
		ext6 := inst.IPv6ExternalInterface
		if ext6 == "" {
			ext6 = ext
		}
		up = append(up,
			fmt.Sprintf("ip6tables -A FORWARD -i %s -j ACCEPT", iface),
			fmt.Sprintf("ip6tables -A FORWARD -o %s -j ACCEPT", iface),
			fmt.Sprintf("ip6tables -A FORWARD -i %s -o %s -j ACCEPT", ext6, iface),
			"sysctl -w net.ipv6.conf.all.forwarding=1",
			fmt.Sprintf("sysctl -w net.ipv6.conf.%s.proxy_ndp=1", ext6),
		)
		down = append(down,
			fmt.Sprintf("ip6tables -D FORWARD -i %s -j ACCEPT", iface),
			fmt.Sprintf("ip6tables -D FORWARD -o %s -j ACCEPT", iface),
			fmt.Sprintf("ip6tables -D FORWARD -i %s -o %s -j ACCEPT", ext6, iface),
		)
		for _, p := range inst.Peers {
			ip6 := firstIPv6(p.AllowedIPs)
			if ip6 == "" {
				continue
			}
			up = append(up, fmt.Sprintf("ip -6 neigh add proxy %s dev %s", ip6, ext6))
			down = append(down, fmt.Sprintf("ip -6 neigh del proxy %s dev %s", ip6, ext6))
		}
	}

	for _, p := range inst.Peers {
		if p.ForwardedPorts == "" {
			continue
		}
		clientIP := FirstIPv4(p.AllowedIPs)
		if clientIP == "" {
			continue
		}
		up = append(up, portForwardLines("-A", ext, iface, clientIP, p.Email, p.ForwardedPorts)...)
		down = append(down, portForwardLines("-D", ext, iface, clientIP, p.Email, p.ForwardedPorts)...)
	}

	if inst.RouteThroughXray {
		egressPort := EgressPortForInbound(inst.Id)
		anyPeerTproxied := false
		for _, p := range inst.Peers {
			clientIP := FirstIPv4(p.AllowedIPs)
			if clientIP == "" {
				continue
			}
			up = append(up, routeEgressLines("-A", iface, clientIP, p.Email, egressPort)...)
			down = append(down, routeEgressLines("-D", iface, clientIP, p.Email, egressPort)...)
			anyPeerTproxied = true
		}
		if anyPeerTproxied {
			// The fwmark->table->local-everywhere policy route is what lets TPROXY
			// deliver a peer's packets to this instance's own Xray bridge even
			// though their destination is never one of this host's own addresses.
			// It is system-wide, not interface-specific, so — like the
			// IPv6-forwarding sysctl above — it is added idempotently here and
			// never torn down in PostDown; a second AmneziaWG instance must find
			// it already in place, not race to remove what the first still needs.
			// "ip rule add" is not itself idempotent (a second call inserts a
			// duplicate rather than deduplicating), and hostRulesFingerprint keys
			// on every peer's presence/IP when RouteThroughXray is on, so PostUp
			// re-runs on any client add/remove/re-IP — without the existence
			// check below, "ip rule show" would accumulate one duplicate entry
			// per bounce forever.
			//
			// TPROXY never rewrites the packet's own destination address — only
			// the routing decision changes, via the fwmark+table trick above — so
			// by the time this packet reaches the host's own INPUT chain, its
			// destination still looks like some remote address (e.g. 8.8.8.8),
			// never this host's own. A default-deny firewall whose INPUT chain
			// sanity-checks "is this destination actually local" (UFW's
			// ufw-not-local, using addrtype --dst-type LOCAL, is exactly this) can
			// never see it as legitimate and silently drops it before Xray's
			// socket ever sees a single byte — TPROXY's own counters keep
			// incrementing the whole time, making this look like a Xray-side bug
			// even though Xray never gets the chance to fail. The fix is the same
			// shape as the policy route above: an idempotent, never-torn-down,
			// system-wide accept for this fwmark, inserted at the very front of
			// the base INPUT chain so it runs before any such sanity check,
			// regardless of which firewall manager (ufw, firewalld, bare
			// iptables) owns the rest of that chain.
			up = append(up,
				// grep -c (not -q): -q exits as soon as it matches, so "ip rule
				// list" can take SIGPIPE; under `set -o pipefail` the pipeline then
				// reports 141 even though the rule WAS found, and "ip rule add"
				// below runs anyway -- reintroducing the exact duplicate-rule
				// accumulation this existence check exists to prevent. -c reads
				// every line to completion and still exits 1 on no match.
				fmt.Sprintf("ip rule list | grep -c 'fwmark %#x lookup %d' >/dev/null || ip rule add fwmark %#x lookup %d", EgressFwmark, EgressTable, EgressFwmark, EgressTable),
				fmt.Sprintf("ip route replace local 0.0.0.0/0 dev lo table %d", EgressTable),
				fmt.Sprintf("iptables -C INPUT -m mark --mark %#x -j ACCEPT 2>/dev/null || iptables -I INPUT 1 -m mark --mark %#x -j ACCEPT", EgressFwmark, EgressFwmark),
			)
		}
	}

	up = append(up, "sysctl -w net.ipv4.ip_forward=1")
	return strings.Join(up, "; "), strings.Join(appendOrTrue(down), "; ")
}

// appendOrTrue suffixes every command with " || true", making the whole
// PostDown chain best-effort. wg-quick/awg-quick joins hook commands with
// "; " and runs the result under `set -e -o pipefail`, so the first non-zero
// command aborts everything after it. On teardown that matters: if
// something has already flushed the filter table out from under the
// interface (a ufw/firewalld reload, fail2ban rebuilding its chains), the
// first "-D" fails and every command after it — including the nat-table DNAT
// deletes a flush does NOT remove — is skipped, and the next PostUp re-adds
// them, accumulating one set per bounce. PostUp is left alone: a real setup
// failure there should still surface, not be silently swallowed.
func appendOrTrue(cmds []string) []string {
	out := make([]string, len(cmds))
	for i, c := range cmds {
		out[i] = c + " || true"
	}
	return out
}

// firstAddress returns the first configured interface address, used as the
// NAT source subnet for PostUp/PostDown.
func firstAddress(addresses []string) string {
	if len(addresses) == 0 {
		return ""
	}
	return addresses[0]
}

// firstIPv6 returns the first IPv6 address (mask stripped) among allowedIPs,
// or "" if none — used to build one NDP proxy PostUp/PostDown entry per peer.
func firstIPv6(allowedIPs []string) string {
	for _, a := range allowedIPs {
		if prefix, err := netip.ParsePrefix(a); err == nil {
			if prefix.Addr().Is6() {
				return prefix.Addr().String()
			}
			continue
		}
		if addr, err := netip.ParseAddr(a); err == nil && addr.Is6() {
			return addr.String()
		}
	}
	return ""
}

// FirstIPv4 returns the first IPv4 address (mask stripped) among allowedIPs,
// or "" if none — used as the DNAT target for a peer's forwarded ports and,
// by internal/web/service's injectAmneziawgEgress, as the source-IP match for
// a routed peer's Xray rule. Exported so both packages derive a peer's
// tunnel IPv4 address the exact same way.
func FirstIPv4(allowedIPs []string) string {
	for _, a := range allowedIPs {
		if prefix, err := netip.ParsePrefix(a); err == nil {
			if prefix.Addr().Is4() {
				return prefix.Addr().String()
			}
			continue
		}
		if addr, err := netip.ParseAddr(a); err == nil && addr.Is4() {
			return addr.String()
		}
	}
	return ""
}

// detectDefaultInterface returns the first non-loopback, non-tunnel, UP
// interface that has a routable IPv4 address. Falls back to "eth0" only if
// nothing is found.
func detectDefaultInterface() string {
	ifaces, err := net.Interfaces()
	if err != nil {
		return "eth0"
	}
	for _, iface := range ifaces {
		if iface.Flags&net.FlagLoopback != 0 || iface.Flags&net.FlagUp == 0 {
			continue
		}
		if strings.HasPrefix(iface.Name, "awg") || strings.HasPrefix(iface.Name, "wg") ||
			strings.HasPrefix(iface.Name, "docker") || strings.HasPrefix(iface.Name, "br-") ||
			strings.HasPrefix(iface.Name, "veth") {
			continue
		}
		addrs, err := iface.Addrs()
		if err != nil || len(addrs) == 0 {
			continue
		}
		for _, addr := range addrs {
			if ipNet, ok := addr.(*net.IPNet); ok && !ipNet.IP.IsLinkLocalUnicast() && ipNet.IP.To4() != nil {
				return iface.Name
			}
		}
	}
	return "eth0"
}

// --- process control ---

func configPath(interfaceName string) string {
	return filepath.Join(configDir, interfaceName+".conf")
}

// writeConfigFile renders and persists the .conf file awg-quick reads.
func writeConfigFile(inst Instance) error {
	if err := os.MkdirAll(configDir, 0o700); err != nil {
		return fmt.Errorf("amneziawg: create config dir: %w", err)
	}
	if err := os.WriteFile(configPath(inst.InterfaceName), []byte(generateServerConfig(inst)), 0o600); err != nil {
		return fmt.Errorf("amneziawg: write config for %s: %w", inst.InterfaceName, err)
	}
	return nil
}

// removeConfigFile deletes the config file for an interface, best-effort.
func removeConfigFile(interfaceName string) {
	if err := os.Remove(configPath(interfaceName)); err != nil && !os.IsNotExist(err) {
		logger.Warningf("amneziawg: failed to remove config file for %s: %v", interfaceName, err)
	}
}

// awgCommandTimeout bounds every short-lived awg/awg-quick invocation so a
// hung command (e.g. a stuck kernel module operation) can't block the
// reconcile job indefinitely.
const awgCommandTimeout = 30 * time.Second

// interfaceUp brings an AmneziaWG interface up via awg-quick.
func interfaceUp(interfaceName string) error {
	ctx, cancel := context.WithTimeout(context.Background(), awgCommandTimeout)
	defer cancel()
	out, err := exec.CommandContext(ctx, "awg-quick", "up", configPath(interfaceName)).CombinedOutput()
	if err != nil {
		return fmt.Errorf("awg-quick up %s failed: %s: %w", interfaceName, strings.TrimSpace(string(out)), err)
	}
	return nil
}

// interfaceDown takes an AmneziaWG interface down via awg-quick.
func interfaceDown(interfaceName string) error {
	ctx, cancel := context.WithTimeout(context.Background(), awgCommandTimeout)
	defer cancel()
	out, err := exec.CommandContext(ctx, "awg-quick", "down", configPath(interfaceName)).CombinedOutput()
	if err != nil {
		return fmt.Errorf("awg-quick down %s failed: %s: %w", interfaceName, strings.TrimSpace(string(out)), err)
	}
	return nil
}

// isInterfaceUp checks whether the named AmneziaWG interface currently
// exists.
func isInterfaceUp(interfaceName string) bool {
	ctx, cancel := context.WithTimeout(context.Background(), awgCommandTimeout)
	defer cancel()
	return exec.CommandContext(ctx, "awg", "show", interfaceName).Run() == nil
}

// syncConfig applies a peers-only config change without dropping existing
// connections on other peers, falling back to a full restart when the live
// interface won't accept the diff (or isn't up yet).
func syncConfig(inst Instance) error {
	if !isInterfaceUp(inst.InterfaceName) {
		return interfaceUp(inst.InterfaceName)
	}

	ctx, cancel := context.WithTimeout(context.Background(), awgCommandTimeout)
	defer cancel()
	stripped, err := exec.CommandContext(ctx, "awg-quick", "strip", configPath(inst.InterfaceName)).Output()
	if err != nil {
		logger.Warningf("amneziawg: awg-quick strip failed for %s, restarting: %v", inst.InterfaceName, err)
		return restartInterface(inst.InterfaceName)
	}

	syncCtx, syncCancel := context.WithTimeout(context.Background(), awgCommandTimeout)
	defer syncCancel()
	sync := exec.CommandContext(syncCtx, "awg", "syncconf", inst.InterfaceName, "/dev/stdin")
	sync.Stdin = bytes.NewReader(stripped)
	if out, err := sync.CombinedOutput(); err != nil {
		logger.Warningf("amneziawg: awg syncconf failed for %s, restarting: %s: %v", inst.InterfaceName, strings.TrimSpace(string(out)), err)
		return restartInterface(inst.InterfaceName)
	}
	return nil
}

// restartInterface performs a full down+up cycle.
func restartInterface(interfaceName string) error {
	_ = interfaceDown(interfaceName)
	return interfaceUp(interfaceName)
}

// peerStat is one peer's runtime stats parsed from `awg show <iface> dump`.
type peerStat struct {
	publicKey       string
	latestHandshake int64 // unix seconds
	rx              int64 // bytes received from the peer (its upload)
	tx              int64 // bytes sent to the peer (its download)
}

// getPeerStats parses `awg show <iface> dump`. The dump format is
// tab-separated: line 1 is the interface (private-key, public-key,
// listen-port, fwmark); each following line is one peer (public-key,
// preshared-key, endpoint, allowed-ips, latest-handshake, transfer-rx,
// transfer-tx, persistent-keepalive).
func getPeerStats(interfaceName string) ([]peerStat, error) {
	ctx, cancel := context.WithTimeout(context.Background(), awgCommandTimeout)
	defer cancel()
	out, err := exec.CommandContext(ctx, "awg", "show", interfaceName, "dump").Output()
	if err != nil {
		return nil, fmt.Errorf("awg show %s dump failed: %w", interfaceName, err)
	}

	var stats []peerStat
	scanner := bufio.NewScanner(bytes.NewReader(out))
	first := true
	for scanner.Scan() {
		if first {
			first = false
			continue
		}
		fields := strings.Split(scanner.Text(), "\t")
		if len(fields) < 8 {
			continue
		}
		handshake, _ := strconv.ParseInt(fields[4], 10, 64)
		rx, _ := strconv.ParseInt(fields[5], 10, 64)
		tx, _ := strconv.ParseInt(fields[6], 10, 64)
		stats = append(stats, peerStat{publicKey: fields[0], latestHandshake: handshake, rx: rx, tx: tx})
	}
	return stats, nil
}

// IsAwgInstalled reports whether the awg and awg-quick binaries are on PATH.
func IsAwgInstalled() bool {
	_, err1 := exec.LookPath("awg")
	_, err2 := exec.LookPath("awg-quick")
	return err1 == nil && err2 == nil
}