Compare commits
2 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
b20ace8763 | ||
|
|
053b38e46c |
@@ -826,12 +826,20 @@ func runSecondaryConfigRender(ctx context.Context, pool *pgxpoolPool, box *secre
|
||||
}
|
||||
|
||||
// runPrimaryPush periodically pushes this secondary node's config_hash to the
|
||||
// primary via mTLS. The primary's ha_nodes view only gets config_hash written
|
||||
// during join-time autoRegister — after that the primary never hears about
|
||||
// hash changes unless we push. Without this, the drift banner shows stale
|
||||
// hashes from join-time forever.
|
||||
// primary via mTLS. The primary's ha_nodes view only gets config_hash + last_seen
|
||||
// written during join-time autoRegister — after that the primary never hears about
|
||||
// the secondary unless we push. Without this, the drift banner shows stale hashes
|
||||
// from join-time forever AND the secondary's last_seen freezes → SweepStaleNodes
|
||||
// marks it offline.
|
||||
//
|
||||
// WICHTIG: tick MUSS deutlich unter dem Stale-Threshold (4× 30s = 2 min, siehe
|
||||
// scheduler.staleThreshold / cluster.SweepStaleNodes) liegen. Sonst flippt der
|
||||
// Secondary zwischen den Pushes zwangsläufig auf "offline" (bei 5-min-Tick:
|
||||
// 2 min online, 3 min offline). 30s = 4 Pushes pro Stale-Fenster → ein
|
||||
// verpasster Push (Netz-Glitch) ist unkritisch. Der Receiver (AgentRegisterPeer)
|
||||
// lädt nftables nur bei IP-Änderung neu → kein Reload-Sturm durch häufige Pushes.
|
||||
func runPrimaryPush(ctx context.Context, pool *pgxpoolPool, nodeID, fqdn, version, primaryURL string) {
|
||||
const tick = 5 * time.Minute
|
||||
const tick = 30 * time.Second
|
||||
t := time.NewTicker(tick)
|
||||
defer t.Stop()
|
||||
push := func() {
|
||||
|
||||
@@ -83,6 +83,8 @@ func main() {
|
||||
slog.Error("waf: SPOE agent stopped", "error", err)
|
||||
os.Exit(1)
|
||||
}
|
||||
// Graceful shutdown (ctx cancelled): gepufferte Alerts flushen.
|
||||
alertWriter.Close()
|
||||
}
|
||||
|
||||
// reload fetches all domain+waf_config pairs from DB and rebuilds engines.
|
||||
|
||||
@@ -7,14 +7,21 @@ import (
|
||||
"net/http"
|
||||
"os"
|
||||
"os/exec"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
"github.com/gin-gonic/gin"
|
||||
|
||||
"git.netcell-it.de/projekte/edgeguard-native/internal/configgen"
|
||||
"git.netcell-it.de/projekte/edgeguard-native/internal/handlers/response"
|
||||
"git.netcell-it.de/projekte/edgeguard-native/internal/models"
|
||||
aptsvc "git.netcell-it.de/projekte/edgeguard-native/internal/services/apt"
|
||||
)
|
||||
|
||||
// ruStateMu serialisiert Lesen/Schreiben der Rolling-Update-State-Datei
|
||||
// (HTTP-Handler + Hintergrund-Goroutine greifen gleichzeitig zu).
|
||||
var ruStateMu sync.Mutex
|
||||
|
||||
const rollingUpdateStateFile = "/var/lib/edgeguard/rolling-update-state.json"
|
||||
|
||||
const (
|
||||
@@ -26,17 +33,24 @@ const (
|
||||
phaseFailed = "failed"
|
||||
)
|
||||
|
||||
// FinishRollingUpdateIfPending wird beim API-Start aufgerufen. Wenn die
|
||||
// State-Datei "updating-primary" enthält, bedeutet das dass der Primary
|
||||
// gerade erfolgreich neugestartet ist → Update abgeschlossen → "done" schreiben.
|
||||
// FinishRollingUpdateIfPending wird beim API-Start aufgerufen.
|
||||
// - "updating-primary": der Primary ist gerade erfolgreich neugestartet →
|
||||
// Update abgeschlossen → "done".
|
||||
// - "updating-secondary"/"waiting-secondary": die orchestrierende Goroutine
|
||||
// lief in DIESEM (jetzt neu gestarteten) Prozess und ist mit ihm gestorben.
|
||||
// Die Phase kann nicht weiterlaufen → auf "idle" zurücksetzen, sonst zeigt
|
||||
// die UI ewig "Rolling Update läuft". (Vorher blieb so ein Stand hängen.)
|
||||
func FinishRollingUpdateIfPending() {
|
||||
st := readRollingUpdateState()
|
||||
if st.Phase == phaseUpdatingPrimary {
|
||||
switch st.Phase {
|
||||
case phaseUpdatingPrimary:
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseDone,
|
||||
SecondaryID: st.SecondaryID,
|
||||
SecondaryFQDN: st.SecondaryFQDN,
|
||||
})
|
||||
case phaseUpdatingSecondary, phaseWaitingSecondary:
|
||||
writeRollingUpdateState(RollingUpdateState{Phase: phaseIdle})
|
||||
}
|
||||
}
|
||||
|
||||
@@ -53,6 +67,8 @@ type RollingUpdateState struct {
|
||||
}
|
||||
|
||||
func readRollingUpdateState() RollingUpdateState {
|
||||
ruStateMu.Lock()
|
||||
defer ruStateMu.Unlock()
|
||||
data, err := os.ReadFile(rollingUpdateStateFile)
|
||||
if err != nil {
|
||||
return RollingUpdateState{Phase: phaseIdle, UpdatedAt: time.Now()}
|
||||
@@ -61,6 +77,13 @@ func readRollingUpdateState() RollingUpdateState {
|
||||
if err := json.Unmarshal(data, &s); err != nil {
|
||||
return RollingUpdateState{Phase: phaseIdle, UpdatedAt: time.Now()}
|
||||
}
|
||||
// Terminale Zustände altern aus (statt Mutation-on-GET): nach 10 min
|
||||
// gilt done/failed als idle — so verliert kein paralleler Poller das
|
||||
// Ergebnis und ein alter Stand bleibt nicht hängen.
|
||||
if (s.Phase == phaseDone || s.Phase == phaseFailed) && !s.UpdatedAt.IsZero() &&
|
||||
time.Since(s.UpdatedAt) > 10*time.Minute {
|
||||
return RollingUpdateState{Phase: phaseIdle, UpdatedAt: time.Now()}
|
||||
}
|
||||
return s
|
||||
}
|
||||
|
||||
@@ -71,7 +94,10 @@ func writeRollingUpdateState(s RollingUpdateState) {
|
||||
slog.Warn("rolling-update: failed to marshal state", "error", err)
|
||||
return
|
||||
}
|
||||
if err := os.WriteFile(rollingUpdateStateFile, data, 0o600); err != nil {
|
||||
ruStateMu.Lock()
|
||||
defer ruStateMu.Unlock()
|
||||
// AtomicWrite (temp+rename) → Leser sehen nie einen partiellen Stand.
|
||||
if err := configgen.AtomicWrite(rollingUpdateStateFile, data, 0o600); err != nil {
|
||||
slog.Warn("rolling-update: failed to write state file", "error", err)
|
||||
}
|
||||
}
|
||||
@@ -127,76 +153,108 @@ func (h *ClusterHandler) RollingUpdate(c *gin.Context) {
|
||||
}
|
||||
|
||||
// RollingUpdateStatus gibt den aktuellen Rolling-Update-State zurück.
|
||||
// Bei phase == "done" wird nach Auslieferung sofort auf idle zurückgesetzt
|
||||
// damit der nächste Pageload keinen Stale-done vorfindet.
|
||||
// Read-only — terminale Zustände altern in readRollingUpdateState aus
|
||||
// (kein Reset-on-GET mehr, das parallelen Pollern das "done" wegnahm).
|
||||
func (h *ClusterHandler) RollingUpdateStatus(c *gin.Context) {
|
||||
st := readRollingUpdateState()
|
||||
response.OK(c, st)
|
||||
if st.Phase == phaseDone {
|
||||
writeRollingUpdateState(RollingUpdateState{Phase: phaseIdle})
|
||||
}
|
||||
response.OK(c, readRollingUpdateState())
|
||||
}
|
||||
|
||||
func (h *ClusterHandler) runRollingUpdate(secondary *models.HANode) {
|
||||
ctx := context.Background()
|
||||
|
||||
// 1. Secondary triggern
|
||||
slog.Info("rolling-update: posting trigger-update to secondary", "fqdn", secondary.FQDN)
|
||||
result := h.Aggregator.PostPeer(ctx, *secondary, "/agent/cluster/trigger-update")
|
||||
if !result.OK {
|
||||
// Zielversion = das verfügbare apt-Candidate (worauf wir hochziehen) und
|
||||
// die aktuelle Secondary-Version als Baseline. Beides steuert, ob der
|
||||
// Secondary überhaupt etwas zu tun hat.
|
||||
candidate := rollingCandidateVersion(ctx)
|
||||
baseline := secondaryVersion(ctx, h, secondary)
|
||||
|
||||
// Ist der Secondary bereits auf der Zielversion, gibt es nichts
|
||||
// hochzuziehen — KEIN Trigger, KEIN Warten. Sonst würde auf einen
|
||||
// Version-Flip gewartet, der nie kommt → 10-min-Timeout (der frühere Bug,
|
||||
// wenn beide Nodes schon aktuell waren).
|
||||
secondaryUpToDate := candidate != "" && baseline != "" && baseline == candidate
|
||||
if secondaryUpToDate {
|
||||
slog.Info("rolling-update: secondary already at target — skipping secondary step",
|
||||
"version", candidate)
|
||||
} else {
|
||||
// 1. Secondary triggern
|
||||
slog.Info("rolling-update: posting trigger-update to secondary", "fqdn", secondary.FQDN)
|
||||
result := h.Aggregator.PostPeer(ctx, *secondary, "/agent/cluster/trigger-update")
|
||||
if !result.OK {
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseFailed,
|
||||
SecondaryID: secondary.ID,
|
||||
SecondaryFQDN: secondary.FQDN,
|
||||
Error: "trigger-update failed: " + result.Err,
|
||||
})
|
||||
slog.Warn("rolling-update: secondary trigger failed", "error", result.Err)
|
||||
return
|
||||
}
|
||||
|
||||
// 2. Secondary-Version pollen — der Secondary restartet nach dem
|
||||
// Upgrade, danach zeigt /agent/cluster/version eine neue Version.
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseFailed,
|
||||
Phase: phaseWaitingSecondary,
|
||||
SecondaryID: secondary.ID,
|
||||
SecondaryFQDN: secondary.FQDN,
|
||||
Error: "trigger-update failed: " + result.Err,
|
||||
})
|
||||
slog.Warn("rolling-update: secondary trigger failed", "error", result.Err)
|
||||
return
|
||||
}
|
||||
slog.Info("rolling-update: waiting for secondary version flip",
|
||||
"baseline", baseline, "candidate", candidate)
|
||||
|
||||
// 2. Secondary-Version pollen — der Secondary restartet nach dem
|
||||
// Upgrade, danach zeigt /agent/cluster/version eine neue Version.
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseWaitingSecondary,
|
||||
SecondaryID: secondary.ID,
|
||||
SecondaryFQDN: secondary.FQDN,
|
||||
})
|
||||
slog.Info("rolling-update: waiting for secondary version flip")
|
||||
// Kurze Wartezeit damit apt auf dem Secondary erst losläuft
|
||||
time.Sleep(20 * time.Second)
|
||||
|
||||
// Kurze Wartezeit damit apt auf dem Secondary erst losläuft
|
||||
time.Sleep(20 * time.Second)
|
||||
|
||||
deadline := time.Now().Add(10 * time.Minute)
|
||||
versionFlipped := false
|
||||
for time.Now().Before(deadline) {
|
||||
results := h.Aggregator.FanOut(ctx, []models.HANode{*secondary}, "/agent/cluster/version", h.LocalID)
|
||||
if len(results) > 0 && results[0].OK {
|
||||
var ver struct {
|
||||
Version string `json:"version"`
|
||||
}
|
||||
if err := json.Unmarshal(results[0].Data, &ver); err == nil {
|
||||
slog.Info("rolling-update: secondary version", "version", ver.Version, "primary", h.Version)
|
||||
if ver.Version != h.Version {
|
||||
versionFlipped = true
|
||||
break
|
||||
deadline := time.Now().Add(10 * time.Minute)
|
||||
versionFlipped := false
|
||||
for time.Now().Before(deadline) {
|
||||
results := h.Aggregator.FanOut(ctx, []models.HANode{*secondary}, "/agent/cluster/version", h.LocalID)
|
||||
if len(results) > 0 && results[0].OK {
|
||||
var ver struct {
|
||||
Version string `json:"version"`
|
||||
}
|
||||
if err := json.Unmarshal(results[0].Data, &ver); err == nil {
|
||||
slog.Info("rolling-update: secondary version", "version", ver.Version,
|
||||
"baseline", baseline, "candidate", candidate)
|
||||
// Erfolg = Secondary hat die Zielversion erreicht (candidate)
|
||||
// ODER hat sich gegenüber der Baseline überhaupt bewegt
|
||||
// (Fallback, wenn candidate nicht ermittelbar war).
|
||||
if ver.Version != "" &&
|
||||
((candidate != "" && ver.Version == candidate) || ver.Version != baseline) {
|
||||
versionFlipped = true
|
||||
break
|
||||
}
|
||||
}
|
||||
}
|
||||
time.Sleep(10 * time.Second)
|
||||
}
|
||||
|
||||
if !versionFlipped {
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseFailed,
|
||||
SecondaryID: secondary.ID,
|
||||
SecondaryFQDN: secondary.FQDN,
|
||||
Error: "timeout (10 min) waiting for secondary version flip",
|
||||
})
|
||||
slog.Warn("rolling-update: secondary version flip timeout")
|
||||
return
|
||||
}
|
||||
time.Sleep(10 * time.Second)
|
||||
}
|
||||
|
||||
if !versionFlipped {
|
||||
// 3. Primary (uns selbst) aktualisieren — identisch zu /system/upgrade.
|
||||
// Ist der Primary bereits auf der Zielversion (z. B. beide Nodes schon
|
||||
// aktuell), gibt es nichts zu tun → direkt "done". Sonst liefe ein
|
||||
// apt-Lauf ohne Paket-Wechsel → kein Restart → Phase hinge ewig in
|
||||
// "updating-primary".
|
||||
if candidate != "" && h.Version == candidate {
|
||||
slog.Info("rolling-update: primary already at target — nothing to upgrade", "version", candidate)
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseFailed,
|
||||
Phase: phaseDone,
|
||||
SecondaryID: secondary.ID,
|
||||
SecondaryFQDN: secondary.FQDN,
|
||||
Error: "timeout (10 min) waiting for secondary version flip",
|
||||
})
|
||||
slog.Warn("rolling-update: secondary version flip timeout")
|
||||
return
|
||||
}
|
||||
|
||||
// 3. Primary (uns selbst) aktualisieren — identisch zu /system/upgrade
|
||||
writeRollingUpdateState(RollingUpdateState{
|
||||
Phase: phaseUpdatingPrimary,
|
||||
SecondaryID: secondary.ID,
|
||||
@@ -256,3 +314,26 @@ rm -f /var/lib/edgeguard/upgrade.sh
|
||||
// UI erkennt Version-Flip via /system/health und schließt den Flow.
|
||||
slog.Info("rolling-update: primary upgrade dispatched, process will restart")
|
||||
}
|
||||
|
||||
// rollingCandidateVersion liefert best-effort die verfügbare apt-Candidate-
|
||||
// Version des Meta-Pakets "edgeguard" — also die Version, auf die das Rolling-
|
||||
// Update hochzieht. Leerer String, wenn apt sie nicht ermitteln kann (dann
|
||||
// fällt runRollingUpdate auf reine Baseline-Flip-Erkennung zurück).
|
||||
func rollingCandidateVersion(ctx context.Context) string {
|
||||
vers := aptsvc.PackageVersions(ctx, false)
|
||||
return vers["edgeguard_available"]
|
||||
}
|
||||
|
||||
// secondaryVersion holt best-effort die laufende Version des Peers via mTLS.
|
||||
func secondaryVersion(ctx context.Context, h *ClusterHandler, secondary *models.HANode) string {
|
||||
results := h.Aggregator.FanOut(ctx, []models.HANode{*secondary}, "/agent/cluster/version", h.LocalID)
|
||||
if len(results) > 0 && results[0].OK {
|
||||
var ver struct {
|
||||
Version string `json:"version"`
|
||||
}
|
||||
if json.Unmarshal(results[0].Data, &ver) == nil {
|
||||
return ver.Version
|
||||
}
|
||||
}
|
||||
return ""
|
||||
}
|
||||
|
||||
@@ -3,6 +3,8 @@ package waf
|
||||
import (
|
||||
"context"
|
||||
"log/slog"
|
||||
"sync"
|
||||
"sync/atomic"
|
||||
"time"
|
||||
|
||||
"github.com/jackc/pgx/v5/pgxpool"
|
||||
@@ -26,8 +28,12 @@ type Alert struct {
|
||||
// AlertWriter accepts Alert values via a buffered channel and writes
|
||||
// them to PostgreSQL asynchronously so SPOE handling stays low-latency.
|
||||
type AlertWriter struct {
|
||||
pool *pgxpool.Pool
|
||||
ch chan Alert
|
||||
pool *pgxpool.Pool
|
||||
ch chan Alert
|
||||
stop chan struct{}
|
||||
done chan struct{}
|
||||
closeOnce sync.Once
|
||||
closed atomic.Bool
|
||||
}
|
||||
|
||||
// NewAlertWriter creates an AlertWriter and starts its background goroutine.
|
||||
@@ -36,14 +42,19 @@ func NewAlertWriter(pool *pgxpool.Pool, bufSize int) *AlertWriter {
|
||||
aw := &AlertWriter{
|
||||
pool: pool,
|
||||
ch: make(chan Alert, bufSize),
|
||||
stop: make(chan struct{}),
|
||||
done: make(chan struct{}),
|
||||
}
|
||||
go aw.run()
|
||||
return aw
|
||||
}
|
||||
|
||||
// Send enqueues an alert. Drops silently if the channel is full to
|
||||
// avoid slowing down SPOE request handling.
|
||||
// Send enqueues an alert. Drops silently if the channel is full (or the
|
||||
// writer is closing) to avoid slowing down / panicking SPOE handling.
|
||||
func (aw *AlertWriter) Send(a Alert) {
|
||||
if aw.closed.Load() {
|
||||
return
|
||||
}
|
||||
select {
|
||||
case aw.ch <- a:
|
||||
default:
|
||||
@@ -51,9 +62,34 @@ func (aw *AlertWriter) Send(a Alert) {
|
||||
}
|
||||
}
|
||||
|
||||
// Close stops the writer and flushes buffered alerts (best-effort).
|
||||
// Safe to call multiple times. The channel is never closed → Send never
|
||||
// panics even if it races with Close.
|
||||
func (aw *AlertWriter) Close() {
|
||||
aw.closeOnce.Do(func() {
|
||||
aw.closed.Store(true)
|
||||
close(aw.stop)
|
||||
})
|
||||
<-aw.done
|
||||
}
|
||||
|
||||
func (aw *AlertWriter) run() {
|
||||
for a := range aw.ch {
|
||||
aw.write(a)
|
||||
defer close(aw.done)
|
||||
for {
|
||||
select {
|
||||
case a := <-aw.ch:
|
||||
aw.write(a)
|
||||
case <-aw.stop:
|
||||
// Restliche gepufferte Alerts noch wegschreiben, dann Ende.
|
||||
for {
|
||||
select {
|
||||
case a := <-aw.ch:
|
||||
aw.write(a)
|
||||
default:
|
||||
return
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
61
internal/waf/alerts_test.go
Normal file
61
internal/waf/alerts_test.go
Normal file
@@ -0,0 +1,61 @@
|
||||
package waf
|
||||
|
||||
import (
|
||||
"context"
|
||||
"os"
|
||||
"sync"
|
||||
"testing"
|
||||
"time"
|
||||
|
||||
"git.netcell-it.de/projekte/edgeguard-native/internal/database"
|
||||
)
|
||||
|
||||
// Beweist Fix #15: AlertWriter.Close() flusht, ist idempotent, und Send/Close
|
||||
// racen ohne Panic (Kanal wird nie geschlossen). Guarded per EG_FWTEST_DSN.
|
||||
func TestAlertWriter_CloseFlush(t *testing.T) {
|
||||
dsn := os.Getenv("EG_FWTEST_DSN")
|
||||
if dsn == "" {
|
||||
t.Skip("set EG_FWTEST_DSN to run the alert-writer test")
|
||||
}
|
||||
ctx := context.Background()
|
||||
var mErr error
|
||||
for i := 0; i < 3; i++ {
|
||||
if mErr = database.Migrate(ctx, dsn); mErr == nil {
|
||||
break
|
||||
}
|
||||
time.Sleep(700 * time.Millisecond)
|
||||
}
|
||||
if mErr != nil {
|
||||
t.Fatalf("migrate: %v", mErr)
|
||||
}
|
||||
pool, err := database.Open(ctx, dsn)
|
||||
if err != nil {
|
||||
t.Fatalf("open: %v", err)
|
||||
}
|
||||
defer pool.Close()
|
||||
|
||||
aw := NewAlertWriter(pool, 64)
|
||||
for i := 0; i < 20; i++ {
|
||||
aw.Send(Alert{Hostname: "t.local", ClientIP: "203.0.113.1", Method: "GET", URI: "/", Action: "detected"})
|
||||
}
|
||||
|
||||
// Send parallel zu Close → darf nicht paniken.
|
||||
var wg sync.WaitGroup
|
||||
for i := 0; i < 10; i++ {
|
||||
wg.Add(1)
|
||||
go func() { defer wg.Done(); aw.Send(Alert{Hostname: "t.local", Action: "detected"}) }()
|
||||
}
|
||||
|
||||
done := make(chan struct{})
|
||||
go func() { aw.Close(); close(done) }()
|
||||
select {
|
||||
case <-done:
|
||||
case <-time.After(10 * time.Second):
|
||||
t.Fatal("Close() did not return (flush hung)")
|
||||
}
|
||||
wg.Wait()
|
||||
|
||||
// Idempotent + Send nach Close ist No-op (kein Panic).
|
||||
aw.Close()
|
||||
aw.Send(Alert{Hostname: "after.local", Action: "detected"})
|
||||
}
|
||||
Reference in New Issue
Block a user