Merge pull request 'Add boot unit applying a local config' (#34) from benvin/boot-unit into main

Reviewed-on: #34

.woodpecker/release.yaml

@@ -59,7 +59,7 @@ steps:
             cpu: 2
 
   - name: package
-    image: git.unkin.net/unkin/almalinux9-rpmbuilder:latest
+    image: artifactapi.k8s.syd1.au.unkin.net/docker-internal/rpmbuilder:0.1.0-alma9
     commands:
       - ./scripts/build-rpm.sh ${CI_COMMIT_TAG}
     depends_on: [build]

DESIGN.md

@@ -430,6 +430,13 @@ report the generation applied, giving a fleet-wide "converged / N behind" view.
   source/dest disables its rule loudly, never opens it.
 - **Adds fail closed, the control plane fails open.** Partial rollout blocks new
   flows until every hop converges; a dead API leaves the last-good posture running.
+- **A generation that severs the API is reverted.** The agent applies as a
+  `tomswall try` does (on-disk snapshot, systemd revert timer, shared lock), then
+  reports `applied` over a fresh connection. If that fails at the transport level,
+  or the apply errors, it records the generation in
+  `/var/lib/tomswall/reverted.json` (skipped until a newer one arrives), restores
+  the snapshot and reports `reverted`/`failed`. A failed restore leaves the timer
+  to retry it and is reported `failed`.
 
 ---
 

cmd/tomswall/agent.go

@@ -29,7 +29,8 @@ func agentCmd() *cobra.Command {
 		Long: `Agent runs the control-plane pull loop: it fetches this device's compiled
 config from tomswallapi, differentially applies it, and reports the applied
 generation back. It caches the last known-good config and, if the control plane
-is unreachable, keeps applying that cache — it never fails closed.
+is unreachable, keeps applying that cache — it never fails closed. A new
+generation that cuts the agent off from the API is reverted and reported as such.
 
 The agent token defaults to the TOMSWALL_AGENT_TOKEN environment variable, and
 the device name defaults to the system hostname.`,

internal/agent/agent.go

@@ -2,8 +2,13 @@ package agent
 
 import (
 	"context"
+	"encoding/json"
+	"errors"
 	"fmt"
 	"log/slog"
+	"net/url"
+	"os"
+	"path/filepath"
 	"time"
 
 	"git.unkin.net/unkin/tomswall/internal/config"
@@ -12,11 +17,17 @@ import (
 )
 
 // Applier applies a translated config to the firewall. Abstracted so the run
-// loop is testable without touching the kernel.
+// loop is testable without touching the kernel. With safe, a change is applied
+// as a pending try: revert restores the previous ruleset (a failed revert leaves
+// the revert timer armed) and keep drops the snapshot. Both are nil when nothing
+// changed or safe is false.
 type Applier interface {
-	Apply(ctx context.Context, cfg *config.Config) error
+	Apply(ctx context.Context, cfg *config.Config, safe bool) (revert, keep func() error, err error)
 }
 
+// revertDelay is when the revert timer fires if the agent dies mid-apply.
+const revertDelay = time.Minute
+
 // Agent runs the pull-apply-report loop for one device.
 type Agent struct {
 	Client   *Client
@@ -25,6 +36,9 @@ type Agent struct {
 	Applier  Applier
 	// Resolver overrides the DNS resolver (tests); nil derives it per-config.
 	Resolver *Resolver
+
+	// lastReverted covers a reverted generation whose persistence failed.
+	lastReverted *reverted
 }
 
 // Run loops until ctx is cancelled, applying one cycle per Interval (and once
@@ -62,16 +76,31 @@ func (a *Agent) RunOnce(ctx context.Context) error {
 			return fmt.Errorf("control plane unreachable and no cached config: %w", err)
 		}
 		// Re-apply last known-good; do not report a generation we didn't fetch.
-		return a.applyConfig(ctx, cached, false)
+		return a.applyConfig(ctx, cached, nil)
 	}
 
-	if err := a.Cache.Write(raw); err != nil {
+	rv, err := a.readReverted()
-		slog.Warn("agent: caching config failed", "err", err)
+	if err != nil {
+		return err
 	}
-	return a.applyConfig(ctx, rc, true)
+	if a.lastReverted != nil && (rv == nil || a.lastReverted.Generation > rv.Generation) {
+		rv = a.lastReverted
+	}
+	if rv != nil {
+		a.reportReverted(ctx, rv)
+		if rc.Generation <= rv.Generation {
+			slog.Info("agent: generation was reverted, waiting for a newer one", "generation", rc.Generation)
+			return nil
+		}
+	}
+	return a.applyConfig(ctx, rc, raw)
 }
 
-func (a *Agent) applyConfig(ctx context.Context, rc *RenderedConfig, report bool) error {
+// applyConfig applies rc. A fetched config (raw != nil) is applied as a pending
+// try, verified by reaching the API through the new ruleset and reverted if that
+// fails; only then is it cached. The cached config is the last verified-good one,
+// so it is applied plainly: there is nothing to verify it against.
+func (a *Agent) applyConfig(ctx context.Context, rc *RenderedConfig, raw []byte) error {
 	resolver := a.Resolver
 	if resolver == nil {
 		resolver = NewResolver(rc.Resolver)
@@ -82,46 +111,219 @@ func (a *Agent) applyConfig(ctx context.Context, rc *RenderedConfig, report bool
 	if err != nil {
 		return fmt.Errorf("translate: %w", err)
 	}
-	if err := a.Applier.Apply(ctx, cfg); err != nil {
+
-		return fmt.Errorf("apply: %w", err)
+	unlock, err := tryapply.Acquire()
+	if errors.Is(err, tryapply.ErrPending) {
+		slog.Warn("agent: a 'tomswall try' is pending, skipping cycle")
+		return nil
+	}
+	if err != nil {
+		return err
+	}
+	defer unlock()
+
+	revert, keep, err := a.Applier.Apply(ctx, cfg, raw != nil)
+	if err != nil {
+		err = fmt.Errorf("apply: %w", err)
+		if raw != nil && revert != nil {
+			return a.revertGeneration(ctx, rc.Generation, StatusFailed, err, revert)
+		}
+		if revert != nil {
+			if rerr := revert(); rerr != nil {
+				err = fmt.Errorf("%w; restore: %v; revert timer pending", err, rerr)
+			}
+		}
+		if raw != nil {
+			if rerr := a.Client.ReportStatus(ctx, Status{Status: StatusFailed, Generation: rc.Generation, Error: err.Error()}); rerr != nil {
+				slog.Warn("agent: reporting status failed", "err", rerr)
+			}
+		}
+		return err
+	}
+	if raw == nil {
+		slog.Info("agent: applied cached config", "generation", rc.Generation, "rules", len(cfg.Rules))
+		return nil
+	}
+
+	if err := a.confirm(ctx, rc.Generation); err != nil {
+		if keep != nil && ctx.Err() == nil {
+			return a.revertGeneration(ctx, rc.Generation, StatusReverted, err, revert)
+		}
+		// Shutdown is not a verdict on the generation: keep it.
+		if keep != nil {
+			if kerr := keep(); kerr != nil {
+				slog.Warn("agent: dropping snapshot failed", "err", kerr)
+			}
+		}
+		return err
+	}
+	if keep != nil {
+		if err := keep(); err != nil {
+			return fmt.Errorf("dropping snapshot: %w", err)
+		}
 	}
 	slog.Info("agent: applied config", "generation", rc.Generation, "rules", len(cfg.Rules))
 
-	if report {
+	if err := a.Cache.Write(raw); err != nil {
-		if err := a.Client.ReportStatus(ctx, rc.Generation); err != nil {
+		slog.Warn("agent: caching config failed", "err", err)
-			slog.Warn("agent: reporting status failed", "err", err)
+	}
-		}
+	a.lastReverted = nil
-		// Report the FIB so the control plane can scope router enforcement.
+	if err := os.Remove(a.revertedPath()); err != nil && !os.IsNotExist(err) {
-		if fib := CollectFIB(ctx); len(fib) > 0 {
+		slog.Warn("agent: clearing reverted generation failed", "err", err)
-			if err := a.Client.ReportRoutes(ctx, fib); err != nil {
+	}
-				slog.Warn("agent: reporting routes failed", "err", err)
+	// Report the FIB so the control plane can scope router enforcement.
-			}
+	if fib := CollectFIB(ctx); len(fib) > 0 {
+		if err := a.Client.ReportRoutes(ctx, fib); err != nil {
+			slog.Warn("agent: reporting routes failed", "err", err)
 		}
 	}
 	return nil
 }
 
-// EngineApplier applies via the real nftables differential engine.
+// revertGeneration marks generation as reverted before restoring, so a failed
-type EngineApplier struct{}
+// restore can never lead to re-applying it, then reports status. It is also kept
+// in memory in case persisting fails. A failed
+// restore leaves the snapshot and timer armed and is reported as failed.
+func (a *Agent) revertGeneration(ctx context.Context, generation int64, status string, cause error, revert func() error) error {
+	rv := &reverted{Generation: generation, Status: status, Error: cause.Error()}
+	a.lastReverted = rv
+	if err := a.writeReverted(rv); err != nil {
+		slog.Error("agent: persisting reverted generation failed", "err", err)
+	}
+	suffix := ""
+	if rerr := revert(); rerr != nil {
+		suffix = fmt.Sprintf("; restore: %v; revert timer pending", rerr)
+		rv.Status = StatusFailed
+		rv.Error += suffix
+		if werr := a.writeReverted(rv); werr != nil {
+			slog.Error("agent: persisting reverted generation failed", "err", werr)
+		}
+	}
+	a.reportReverted(ctx, rv)
+	return fmt.Errorf("generation %d %s: %w%s", generation, rv.Status, cause, suffix)
+}
 
-// Apply computes and applies the differential change set for cfg. It refuses
+var (
-// while a 'tomswall try' awaits confirmation.
+	verifyAttempts = 3
-func (EngineApplier) Apply(_ context.Context, cfg *config.Config) error {
+	verifyDelay    = 2 * time.Second
-	unlock, err := tryapply.Acquire()
+	verifyTimeout  = 5 * time.Second
+)
+
+// errUnreachable means the API could not be reached through the new ruleset.
+var errUnreachable = errors.New("control plane unreachable after apply")
+
+// confirm reports generation as applied over a fresh connection, which proves
+// the API is reachable through the new ruleset. Any HTTP response counts as
+// reachable; only repeated transport failures return errUnreachable.
+func (a *Agent) confirm(ctx context.Context, generation int64) error {
+	var err error
+	for i := 0; i < verifyAttempts; i++ {
+		if i > 0 {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case <-time.After(verifyDelay):
+			}
+		}
+		actx, cancel := context.WithTimeout(ctx, verifyTimeout)
+		err = a.Client.ReportStatus(actx, Status{Status: StatusApplied, Generation: generation})
+		cancel()
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+		var uerr *url.Error
+		if !errors.As(err, &uerr) {
+			if err != nil {
+				slog.Warn("agent: reporting status failed", "err", err)
+			}
+			return nil
+		}
+	}
+	return fmt.Errorf("%w: %v", errUnreachable, err)
+}
+
+// reverted is a generation rolled back after a failed apply or for severing the
+// API; persisted so it is not re-applied until a newer generation arrives.
+type reverted struct {
+	Generation int64  `json:"generation"`
+	Status     string `json:"status,omitempty"`
+	Error      string `json:"error,omitempty"`
+	Reported   bool   `json:"reported"`
+}
+
+func (a *Agent) revertedPath() string {
+	return filepath.Join(filepath.Dir(a.Cache.Path), "reverted.json")
+}
+
+func (a *Agent) readReverted() (*reverted, error) {
+	b, err := os.ReadFile(a.revertedPath())
+	if os.IsNotExist(err) {
+		return nil, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	var rv reverted
+	if err := json.Unmarshal(b, &rv); err != nil {
+		return nil, fmt.Errorf("parsing %s: %w", a.revertedPath(), err)
+	}
+	return &rv, nil
+}
+
+func (a *Agent) writeReverted(rv *reverted) error {
+	b, err := json.Marshal(rv)
 	if err != nil {
 		return err
 	}
-	defer unlock()
+	return tryapply.WriteFile(a.revertedPath(), b)
+}
+
+// reportReverted reports rv until the control plane accepts it.
+func (a *Agent) reportReverted(ctx context.Context, rv *reverted) {
+	if rv.Reported {
+		return
+	}
+	status := rv.Status
+	if status == "" {
+		status = StatusReverted
+	}
+	if err := a.Client.ReportStatus(ctx, Status{Status: status, Generation: rv.Generation, Error: rv.Error}); err != nil {
+		slog.Warn("agent: reporting reverted generation failed, retrying next cycle", "generation", rv.Generation, "err", err)
+		return
+	}
+	rv.Reported = true
+	if err := a.writeReverted(rv); err != nil {
+		slog.Warn("agent: persisting reverted generation failed", "err", err)
+	}
+}
+
+// EngineApplier applies via the real nftables differential engine.
+type EngineApplier struct{}
+
+// Apply computes and applies the differential change set for cfg, with safe
+// under a pending try as 'tomswall try' does. The caller holds the try lock.
+func (EngineApplier) Apply(_ context.Context, cfg *config.Config, safe bool) (revert, keep func() error, err error) {
 	engine, err := nftables.NewEngine(cfg)
 	if err != nil {
-		return fmt.Errorf("initializing nftables: %w", err)
+		return nil, nil, fmt.Errorf("initializing nftables: %w", err)
 	}
 	changes, err := engine.Plan()
 	if err != nil {
-		return fmt.Errorf("computing changes: %w", err)
+		return nil, nil, fmt.Errorf("computing changes: %w", err)
 	}
 	if changes.Empty() {
-		return nil
+		return nil, nil, nil
 	}
-	return engine.Apply(changes)
+	if !safe {
+		return nil, nil, engine.Apply(changes)
+	}
+	snap, err := engine.Snapshot()
+	if err != nil {
+		return nil, nil, fmt.Errorf("snapshotting ruleset: %w", err)
+	}
+	// PID 0: 'tomswall confirm' must not signal the agent.
+	if _, err := tryapply.Arm(snap, 0, revertDelay); err != nil {
+		return nil, nil, err
+	}
+	return tryapply.Abort, tryapply.Discard, engine.Apply(changes)
 }

internal/agent/agent_test.go

@@ -132,10 +132,10 @@ type fakeApplier struct {
 	lastGen int
 }
 
-func (f *fakeApplier) Apply(_ context.Context, cfg *config.Config) error {
+func (f *fakeApplier) Apply(_ context.Context, cfg *config.Config, _ bool) (func() error, func() error, error) {
 	atomic.AddInt32(&f.count, 1)
 	f.lastGen = len(cfg.Rules)
-	return nil
+	return nil, nil, nil
 }
 
 const renderedYAML = `generation: 7

internal/agent/cache.go

@@ -2,7 +2,8 @@ package agent
 
 import (
 	"os"
-	"path/filepath"
+
+	"git.unkin.net/unkin/tomswall/internal/tryapply"
 )
 
 // Cache persists the last known-good rendered config to disk so the agent can
@@ -11,16 +12,9 @@ type Cache struct {
 	Path string
 }
 
-// Write atomically stores the raw config bytes.
+// Write durably stores the raw config bytes.
 func (c Cache) Write(raw []byte) error {
-	if err := os.MkdirAll(filepath.Dir(c.Path), 0o755); err != nil {
+	return tryapply.WriteFile(c.Path, raw)
-		return err
-	}
-	tmp := c.Path + ".tmp"
-	if err := os.WriteFile(tmp, raw, 0o600); err != nil {
-		return err
-	}
-	return os.Rename(tmp, c.Path)
 }
 
 // Read returns the cached config, or (nil, nil) when no cache exists yet.

internal/agent/client.go

@@ -26,7 +26,9 @@ func NewClient(baseURL, device, token string) *Client {
 		BaseURL: baseURL,
 		Device:  device,
 		Token:   token,
-		HTTP:    &http.Client{Timeout: 30 * time.Second},
+		// No keep-alives: every request, the post-apply check included, opens a
+		// fresh connection that must pass the current ruleset.
+		HTTP: &http.Client{Timeout: 30 * time.Second, Transport: noKeepAlive()},
 	}
 }
 
@@ -94,10 +96,24 @@ func (c *Client) ReportRoutes(ctx context.Context, prefixes []string) error {
 	return nil
 }
 
-// ReportStatus tells the control plane which generation this device has applied.
+// Status values reported to POST /api/v1/devices/{name}/status.
-func (c *Client) ReportStatus(ctx context.Context, generation int64) error {
+const (
+	StatusApplied  = "applied"
+	StatusReverted = "reverted"
+	StatusFailed   = "failed"
+)
+
+// Status is the outcome of applying one generation.
+type Status struct {
+	Status     string `json:"status"`
+	Generation int64  `json:"generation"`
+	Error      string `json:"error,omitempty"`
+}
+
+// ReportStatus tells the control plane the outcome of applying a generation.
+func (c *Client) ReportStatus(ctx context.Context, st Status) error {
 	url := fmt.Sprintf("%s/api/v1/devices/%s/status", c.BaseURL, c.Device)
-	payload, _ := json.Marshal(map[string]int64{"generation": generation})
+	payload, _ := json.Marshal(st)
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(payload))
 	if err != nil {
 		return err
@@ -116,3 +132,9 @@ func (c *Client) ReportStatus(ctx context.Context, generation int64) error {
 	}
 	return nil
 }
+
+func noKeepAlive() http.RoundTripper {
+	t := http.DefaultTransport.(*http.Transport).Clone()
+	t.DisableKeepAlives = true
+	return t
+}

internal/agent/safeapply_test.go

@@ -0,0 +1,396 @@
+package agent
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"git.unkin.net/unkin/tomswall/internal/config"
+	"git.unkin.net/unkin/tomswall/internal/nftables"
+	"git.unkin.net/unkin/tomswall/internal/tryapply"
+)
+
+func TestMain(m *testing.M) {
+	dir, err := os.MkdirTemp("", "tomswall-agent-test")
+	if err != nil {
+		panic(err)
+	}
+	tryapply.Dir = dir
+	tryapply.Run = func(name string, args ...string) error {
+		timerCmds = append(timerCmds, name)
+		return nil
+	}
+	verifyDelay = time.Millisecond
+	verifyTimeout = time.Second
+	code := m.Run()
+	os.RemoveAll(dir)
+	os.Exit(code)
+}
+
+// fakeAPI serves a config generation and records status reports; while cut it
+// drops connections to the status endpoint, as a severing ruleset would.
+type fakeAPI struct {
+	*httptest.Server
+	gen     atomic.Int64
+	cut     atomic.Bool
+	code    atomic.Int32
+	mu      sync.Mutex
+	reports []Status
+}
+
+func newFakeAPI(t *testing.T, gen int64) *fakeAPI {
+	f := &fakeAPI{}
+	f.gen.Store(gen)
+	f.code.Store(http.StatusNoContent)
+	f.Server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/api/v1/devices/fw-a/config":
+			_, _ = w.Write([]byte(strings.Replace(renderedYAML, "generation: 7", "generation: "+itoa(f.gen.Load()), 1)))
+		case "/api/v1/devices/fw-a/status":
+			if f.cut.Load() {
+				conn, _, _ := w.(http.Hijacker).Hijack()
+				conn.Close()
+				return
+			}
+			var st Status
+			_ = json.NewDecoder(r.Body).Decode(&st)
+			f.mu.Lock()
+			f.reports = append(f.reports, st)
+			f.mu.Unlock()
+			w.WriteHeader(int(f.code.Load()))
+		default:
+			w.WriteHeader(http.StatusNotFound)
+		}
+	}))
+	t.Cleanup(f.Close)
+	return f
+}
+
+// timerCmds records the systemd commands tryapply runs.
+var timerCmds []string
+
+func itoa(n int64) string { b, _ := json.Marshal(n); return string(b) }
+
+func (f *fakeAPI) last() Status {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if len(f.reports) == 0 {
+		return Status{}
+	}
+	return f.reports[len(f.reports)-1]
+}
+
+// fakeEngine always changes the ruleset, when safe under a real tryapply pending
+// try; onApply simulates its effect and restoreErr fails the restore.
+type fakeEngine struct {
+	applies, plain, restores int
+	err, restoreErr          error
+	onApply                  func()
+	onRestore                func()
+}
+
+func (f *fakeEngine) Apply(_ context.Context, _ *config.Config, safe bool) (func() error, func() error, error) {
+	if !safe {
+		f.plain++
+		return nil, nil, f.err
+	}
+	if _, err := tryapply.Arm(&nftables.Snapshot{Table: "tomswall"}, 0, time.Minute); err != nil {
+		return nil, nil, err
+	}
+	tryapply.Restore = func(*nftables.Snapshot) error {
+		f.restores++
+		if f.onRestore != nil {
+			f.onRestore()
+		}
+		return f.restoreErr
+	}
+	f.applies++
+	if f.onApply != nil {
+		f.onApply()
+	}
+	return tryapply.Abort, tryapply.Discard, f.err
+}
+
+// pending reports whether a snapshot is still armed and its timer not stopped since.
+func pending(t *testing.T) bool {
+	t.Helper()
+	_, err := os.Stat(filepath.Join(tryapply.Dir, "try-snapshot.json"))
+	armed := len(timerCmds) > 0 && timerCmds[len(timerCmds)-1] == "systemd-run"
+	if (err == nil) != armed {
+		t.Fatalf("snapshot present=%v but timer armed=%v", err == nil, armed)
+	}
+	return armed
+}
+
+func newAgent(t *testing.T, api *fakeAPI, eng *fakeEngine) *Agent {
+	return &Agent{
+		Client:  NewClient(api.URL, "fw-a", "tok"),
+		Cache:   Cache{Path: filepath.Join(t.TempDir(), "rendered.yaml")},
+		Applier: eng,
+	}
+}
+
+func cachedGen(t *testing.T, a *Agent) int64 {
+	rc, err := a.Cache.Read()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if rc == nil {
+		return 0
+	}
+	return rc.Generation
+}
+
+func TestSafeApplyReachableApplies(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.restores != 0 || api.last() != (Status{Status: StatusApplied, Generation: 7}) || cachedGen(t, a) != 7 || pending(t) {
+		t.Fatalf("restores=%d last=%+v cache=%d", eng.restores, api.last(), cachedGen(t, a))
+	}
+}
+
+func TestSafeApplyUnreachableRevertsAndReports(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{onApply: func() { api.cut.Store(true) }, onRestore: func() { api.cut.Store(false) }}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); !errors.Is(err, errUnreachable) {
+		t.Fatalf("want errUnreachable, got %v", err)
+	}
+	if eng.restores != 1 || cachedGen(t, a) != 0 || pending(t) {
+		t.Fatalf("restores=%d cache=%d", eng.restores, cachedGen(t, a))
+	}
+	if st := api.last(); st.Status != StatusReverted || st.Generation != 7 || st.Error == "" {
+		t.Fatalf("last report %+v", st)
+	}
+	rv, _ := a.readReverted()
+	if rv == nil || rv.Generation != 7 || !rv.Reported {
+		t.Fatalf("persisted %+v", rv)
+	}
+}
+
+func TestSafeApplyRevertReportedOnceReachable(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{onApply: func() { api.cut.Store(true) }}
+	a := newAgent(t, api, eng)
+	_ = a.RunOnce(context.Background())
+	if eng.restores != 1 || api.last().Status != "" {
+		t.Fatalf("restores=%d last=%+v", eng.restores, api.last())
+	}
+	api.cut.Store(false)
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.applies != 1 || api.last() != (Status{Status: StatusReverted, Generation: 7, Error: api.last().Error}) {
+		t.Fatalf("applies=%d last=%+v", eng.applies, api.last())
+	}
+}
+
+func TestSafeApplyShutdownDoesNotRevert(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	ctx, cancel := context.WithCancel(context.Background())
+	eng := &fakeEngine{onApply: func() { api.cut.Store(true); cancel() }}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(ctx); !errors.Is(err, context.Canceled) {
+		t.Fatalf("want context.Canceled, got %v", err)
+	}
+	if rv, _ := a.readReverted(); eng.restores != 0 || rv != nil || cachedGen(t, a) != 0 {
+		t.Fatalf("restores=%d reverted=%+v cache=%d", eng.restores, rv, cachedGen(t, a))
+	}
+}
+
+func TestSafeApplyHTTPErrorDoesNotRevert(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	api.code.Store(http.StatusInternalServerError)
+	eng := &fakeEngine{}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.restores != 0 || cachedGen(t, a) != 7 {
+		t.Fatalf("restores=%d cache=%d", eng.restores, cachedGen(t, a))
+	}
+}
+
+func TestSafeApplyApplyErrorRestoresAndReportsFailed(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{err: errors.New("netlink: boom")}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); err == nil {
+		t.Fatal("want error")
+	}
+	if st := api.last(); eng.restores != 1 || st.Status != StatusFailed || !strings.Contains(st.Error, "boom") || pending(t) {
+		t.Fatalf("restores=%d last=%+v", eng.restores, st)
+	}
+	if rv, _ := a.readReverted(); rv == nil || rv.Generation != 7 || !rv.Reported {
+		t.Fatalf("persisted %+v", rv)
+	}
+}
+
+func TestSafeApplyApplyErrorRestoreFailsKeepsTimer(t *testing.T) {
+	t.Cleanup(func() { _ = tryapply.Discard() })
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{err: errors.New("netlink: boom"), restoreErr: errors.New("netlink: stuck")}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); err == nil || !strings.Contains(err.Error(), "restore: restoring snapshot: netlink: stuck") {
+		t.Fatalf("got %v", err)
+	}
+	want := "apply: netlink: boom; restore: restoring snapshot: netlink: stuck; revert timer pending"
+	if st := api.last(); st != (Status{Status: StatusFailed, Generation: 7, Error: want}) || !pending(t) {
+		t.Fatalf("last=%+v", st)
+	}
+	if rv, _ := a.readReverted(); rv == nil || rv.Generation != 7 || rv.Status != StatusFailed {
+		t.Fatalf("persisted %+v", rv)
+	}
+	// The next cycle waits for the timer instead of re-applying.
+	if err := a.RunOnce(context.Background()); err != nil || eng.applies != 1 {
+		t.Fatalf("err=%v applies=%d", err, eng.applies)
+	}
+}
+
+func TestSafeApplyUnreachableRestoreFailsKeepsTimer(t *testing.T) {
+	t.Cleanup(func() { _ = tryapply.Discard() })
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{onApply: func() { api.cut.Store(true) }, restoreErr: errors.New("netlink: stuck")}
+	eng.onRestore = func() { api.cut.Store(false) }
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); !errors.Is(err, errUnreachable) || !strings.Contains(err.Error(), "revert timer pending") {
+		t.Fatalf("got %v", err)
+	}
+	st := api.last()
+	if st.Status != StatusFailed || st.Generation != 7 || !strings.HasPrefix(st.Error, errUnreachable.Error()) ||
+		!strings.HasSuffix(st.Error, "; restore: restoring snapshot: netlink: stuck; revert timer pending") || !pending(t) {
+		t.Fatalf("last=%+v", st)
+	}
+	if rv, _ := a.readReverted(); rv == nil || rv.Generation != 7 || !rv.Reported || cachedGen(t, a) != 0 {
+		t.Fatalf("persisted %+v cache=%d", rv, cachedGen(t, a))
+	}
+}
+
+func TestSafeApplyRevertedGenerationSkippedAfterRestart(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{}
+	a := newAgent(t, api, eng)
+	if err := a.writeReverted(&reverted{Generation: 7, Reported: true}); err != nil {
+		t.Fatal(err)
+	}
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.applies != 0 {
+		t.Fatalf("reverted generation re-applied")
+	}
+
+	api.gen.Store(8)
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if rv, _ := a.readReverted(); eng.applies != 1 || api.last().Generation != 8 || rv != nil {
+		t.Fatalf("applies=%d last=%+v reverted=%+v", eng.applies, api.last(), rv)
+	}
+}
+
+func TestSafeApplySkipsWhileTryPending(t *testing.T) {
+	marker := filepath.Join(tryapply.Dir, "try-snapshot.json")
+	if err := os.WriteFile(marker, []byte("{}"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	defer os.Remove(marker)
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.applies != 0 || api.last().Status != "" {
+		t.Fatalf("applies=%d last=%+v", eng.applies, api.last())
+	}
+}
+
+// failArm makes arming the revert timer fail, as without systemd.
+func failArm(t *testing.T) {
+	orig := tryapply.Run
+	tryapply.Run = func(name string, args ...string) error {
+		timerCmds = append(timerCmds, name)
+		if name == "systemd-run" {
+			return errors.New("no systemd")
+		}
+		return nil
+	}
+	t.Cleanup(func() { tryapply.Run = orig })
+}
+
+func TestSafeApplyArmFailureReportsFailedAndRetries(t *testing.T) {
+	failArm(t)
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{}
+	a := newAgent(t, api, eng)
+	if err := a.RunOnce(context.Background()); err == nil || !strings.Contains(err.Error(), "no systemd") {
+		t.Fatalf("got %v", err)
+	}
+	if st := api.last(); eng.applies != 0 || st.Status != StatusFailed || st.Generation != 7 || pending(t) || cachedGen(t, a) != 0 {
+		t.Fatalf("applies=%d last=%+v", eng.applies, st)
+	}
+	if rv, _ := a.readReverted(); rv != nil || a.lastReverted != nil {
+		t.Fatalf("arm failure marked generation reverted: %+v", rv)
+	}
+	tryapply.Run = func(name string, args ...string) error {
+		timerCmds = append(timerCmds, name)
+		return nil
+	}
+	if err := a.RunOnce(context.Background()); err != nil || eng.applies != 1 || cachedGen(t, a) != 7 {
+		t.Fatalf("retry err=%v applies=%d", err, eng.applies)
+	}
+}
+
+func TestCachedConfigAppliesWithoutArm(t *testing.T) {
+	failArm(t)
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{}
+	a := newAgent(t, api, eng)
+	if err := a.Cache.Write([]byte(renderedYAML)); err != nil {
+		t.Fatal(err)
+	}
+	api.Close()
+	if err := a.RunOnce(context.Background()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.plain != 1 || eng.applies != 0 || pending(t) {
+		t.Fatalf("plain=%d safe=%d", eng.plain, eng.applies)
+	}
+}
+
+func TestSafeApplyRevertedKeptInMemoryWhenPersistFails(t *testing.T) {
+	api := newFakeAPI(t, 7)
+	eng := &fakeEngine{onRestore: func() { api.cut.Store(false) }}
+	a := newAgent(t, api, eng)
+	// A non-empty directory in its place makes persisting reverted.json fail.
+	eng.onApply = func() {
+		api.cut.Store(true)
+		_ = os.MkdirAll(filepath.Join(a.revertedPath(), "x"), 0o755)
+	}
+	if err := a.RunOnce(context.Background()); !errors.Is(err, errUnreachable) {
+		t.Fatalf("want errUnreachable, got %v", err)
+	}
+	if err := os.RemoveAll(a.revertedPath()); err != nil {
+		t.Fatal(err)
+	}
+	if eng.restores != 1 || api.last().Status != StatusReverted || pending(t) {
+		t.Fatalf("restores=%d last=%+v", eng.restores, api.last())
+	}
+	if err := a.RunOnce(context.Background()); err != nil || eng.applies != 1 {
+		t.Fatalf("reverted generation re-applied: err=%v applies=%d", err, eng.applies)
+	}
+}

internal/config/config.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"regexp"
 	"strings"
 
 	"gopkg.in/yaml.v3"
@@ -55,7 +56,9 @@ type Settings struct {
 	AddressFamily AddressFamily `yaml:"address_family,omitempty"`
 	IPForwarding  bool          `yaml:"ip_forwarding"`
 	LogLevel      string        `yaml:"log_level"`
-	TableName     string        `yaml:"table_name"`
+	// LogLimit rate-limits every log site, shorewall LOGLIMIT syntax rate/unit[:burst]; unset logs every hit.
+	LogLimit  string `yaml:"log_limit,omitempty"`
+	TableName string `yaml:"table_name"`
 
 	// When true, auto-generate CONTINUE policies for sub-zones to their parent zones.
 	ImplicitContinue bool `yaml:"implicit_continue,omitempty"`
@@ -112,6 +115,8 @@ func (c *Config) applyDefaults() {
 	}
 }
 
+var logLimitRe = regexp.MustCompile(`^[1-9][0-9]*/(sec|second|min|minute|hour|day)(:[1-9][0-9]*)?$`)
+
 var validAddressFamilies = map[AddressFamily]bool{
 	FamilyINET: true, FamilyIP: true, FamilyIP6: true,
 }
@@ -120,6 +125,9 @@ func (c *Config) validateSettings() error {
 	if !validAddressFamilies[c.Settings.AddressFamily] {
 		return fmt.Errorf("unknown address_family %q (use inet, ip, or ip6)", c.Settings.AddressFamily)
 	}
+	if l := c.Settings.LogLimit; l != "" && !logLimitRe.MatchString(l) {
+		return fmt.Errorf("invalid log_limit %q (use rate/{sec|min|hour|day}[:burst]; per-source s:/d: is not supported)", l)
+	}
 	for name, d := range map[string]PolicyAction{
 		"invalid_disposition":   c.Settings.InvalidDisposition,
 		"untracked_disposition": c.Settings.UntrackedDisposition,

internal/config/config_test.go

@@ -1069,3 +1069,15 @@ func TestValidateDispositions(t *testing.T) {
 		checkErr(t, c.Validate(), tc.wantErr)
 	}
 }
+
+func TestValidateLogLimit(t *testing.T) {
+	for v, ok := range map[string]bool{
+		"": true, "1/sec": true, "1/sec:10": true, "30/minute:5": true, "2/hour": true, "1/day:1": true,
+		"s:1/sec:10": false, "d:1/sec": false, "1": false, "1/week": false, "0/sec": false, "1/sec:": false,
+	} {
+		c := &Config{Settings: Settings{AddressFamily: FamilyINET, LogLimit: v}}
+		if err := c.validateSettings(); (err == nil) != ok {
+			t.Errorf("log_limit %q: err = %v, want ok=%v", v, err, ok)
+		}
+	}
+}

internal/nftables/compiler.go

@@ -65,10 +65,39 @@ func (c *Compiler) Compile() (*FirewallState, error) {
 		return nil, fmt.Errorf("static-nat: %w", err)
 	}
 	c.compileMSSClamp(state)
+	limitLogs(state, c.cfg.Settings.LogLimit)
 
 	return state, nil
 }
 
+// limitLogs puts a limit in front of every log expression. A limit stops the
+// whole rule, so like shorewall's separate LOG rule, a log followed by an action
+// splits into a limited log-only rule and the same rule without the log. A
+// LOG-action rule (nothing but rate limits after the log) keeps only the log rule.
+func limitLogs(state *FirewallState, spec string) {
+	if spec == "" {
+		return
+	}
+	for chain, rules := range state.Rules {
+		var out []ManagedRule
+		for _, r := range rules {
+			i := slices.IndexFunc(r.Exprs, func(e expr.Any) bool { _, ok := e.(*expr.Log); return ok })
+			if i < 0 {
+				out = append(out, r)
+				continue
+			}
+			logRule := r
+			logRule.Exprs = slices.Concat(r.Exprs[:i], parseRateLimit(spec), r.Exprs[i:i+1])
+			out = append(out, logRule)
+			if slices.ContainsFunc(r.Exprs[i+1:], func(e expr.Any) bool { _, ok := e.(*expr.Limit); return !ok }) {
+				r.Exprs = slices.Concat(r.Exprs[:i], r.Exprs[i+1:])
+				out = append(out, r)
+			}
+		}
+		state.Rules[chain] = out
+	}
+}
+
 func (c *Compiler) compileConntrackFastPath(state *FirewallState) error {
 	invalid := c.cfg.Settings.InvalidDisposition
 	if invalid == "" {
@@ -414,24 +443,24 @@ func (c *Compiler) applyChainExtras(state *FirewallState, chain, tag string, rul
 			break
 		}
 
-		var extra []expr.Any
+		var match, extra []expr.Any
 		var replaceVerdict []expr.Any
 
 		if rule.User != "" {
-			extra = append(extra, matchUID(rule.User)...)
+			match = append(match, matchUID(rule.User)...)
 		}
 		if rule.Mark != "" {
-			extra = append(extra, matchMark(rule.Mark)...)
+			match = append(match, matchMark(rule.Mark)...)
+		}
+		if rule.ConnLimit != "" {
+			match = append(match, matchConnLimit(rule.ConnLimit)...)
+		}
+		if rule.Time != nil {
+			match = append(match, matchTime(rule.Time)...)
 		}
 		if rule.RateLimit != "" {
 			extra = append(extra, parseRateLimit(rule.RateLimit)...)
 		}
-		if rule.ConnLimit != "" {
-			extra = append(extra, matchConnLimit(rule.ConnLimit)...)
-		}
-		if rule.Time != nil {
-			extra = append(extra, matchTime(rule.Time)...)
-		}
 		if rule.SetMark != "" {
 			extra = append(extra, setMarkExprs(rule.SetMark)...)
 		}
@@ -439,21 +468,27 @@ func (c *Compiler) applyChainExtras(state *FirewallState, chain, tag string, rul
 			replaceVerdict = []expr.Any{&expr.Queue{Num: uint16(rule.NFQueue), Total: 1}}
 		}
 
-		if len(extra) > 0 || len(replaceVerdict) > 0 {
+		if len(match) > 0 || len(extra) > 0 || len(replaceVerdict) > 0 {
-			existingExprs := rules[idx].Exprs
+			var pre, post, verdict []expr.Any
-			var verdict []expr.Any
+			for _, e := range rules[idx].Exprs {
-			var nonVerdict []expr.Any
+				switch e.(type) {
-			for _, e := range existingExprs {
+				case *expr.Verdict:
-				if _, ok := e.(*expr.Verdict); ok {
 					verdict = append(verdict, e)
-				} else {
+				case *expr.Log:
-					nonVerdict = append(nonVerdict, e)
+					post = append(post, e)
+				default:
+					if len(post) > 0 {
+						post = append(post, e)
+					} else {
+						pre = append(pre, e)
+					}
 				}
 			}
 			if len(replaceVerdict) > 0 {
 				verdict = replaceVerdict
 			}
-			rules[idx].Exprs = append(append(nonVerdict, extra...), verdict...)
+			// matches go before the log so it only fires for packets the rule matches
+			rules[idx].Exprs = slices.Concat(pre, match, post, extra, verdict)
 		}
 	}
 	state.Rules[chain] = rules

internal/nftables/compiler_test.go

@@ -3155,3 +3155,168 @@ func TestCompile_Dispositions(t *testing.T) {
 		}
 	}
 }
+
+func TestCompile_LogLimit(t *testing.T) {
+	cfg := &config.Config{
+		Settings: config.Settings{
+			TableName:     "test",
+			AddressFamily: config.FamilyINET,
+			LogLimit:      "1/sec:10",
+		},
+		Zones: map[string]config.Zone{
+			"fw":  {Type: config.ZoneFirewall},
+			"net": {Type: config.ZoneIP},
+		},
+		Interfaces: []config.Interface{{Zone: "net", Interface: "eth0"}},
+		Policy: []config.Policy{
+			{Source: "net", Dest: "all", Action: config.PolicyDrop, Log: "info"},
+		},
+		Rules: []config.Rule{
+			{Action: config.RuleAccept, Source: "net:192.0.2.1", Dest: "fw", Proto: "tcp", DPort: config.PortSpec{"22"}, Log: "info"},
+			{Action: config.RuleLog, Source: "net:198.51.100.0/24", Dest: "fw", Proto: "tcp", DPort: config.PortSpec{"23"}, Log: "info"},
+		},
+		PortGroups: make(map[string]config.PortGroup),
+	}
+	state, err := NewCompiler(cfg).Compile()
+	if err != nil {
+		t.Fatalf("Compile() error: %v", err)
+	}
+
+	want := &expr.Limit{Type: expr.LimitTypePkts, Rate: 1, Unit: expr.LimitTimeSecond, Burst: 10}
+	byTag := map[string][]ManagedRule{}
+	for _, r := range state.Rules["input"] {
+		byTag[r.Tag] = append(byTag[r.Tag], r)
+	}
+	for tag, n := range map[string]int{"policy:0": 2, "rule:0": 2, "rule:1": 1} {
+		rs := byTag[tag]
+		if len(rs) != n {
+			t.Fatalf("%s: %d rules, want %d", tag, len(rs), n)
+		}
+		logRule := rs[0].Exprs
+		l := len(logRule)
+		if l < 2 || !reflect.DeepEqual(logRule[l-2], want) {
+			t.Errorf("%s: want limit before log, got %#v", tag, logRule)
+		}
+		if _, ok := logRule[l-1].(*expr.Log); !ok {
+			t.Errorf("%s: log rule must end in log, got %T", tag, logRule[l-1])
+		}
+		for _, e := range rs[n-1].Exprs[:len(rs[n-1].Exprs)-1] {
+			if _, ok := e.(*expr.Log); n == 2 && ok {
+				t.Errorf("%s: verdict rule still logs", tag)
+			}
+		}
+	}
+	if logs := byTag["policy:0"]; len(logs) == 2 {
+		if _, ok := logs[1].Exprs[len(logs[1].Exprs)-1].(*expr.Verdict); !ok {
+			t.Errorf("policy verdict rule must end in a verdict")
+		}
+	}
+}
+
+func TestCompile_NoLogLimitKeepsInlineLog(t *testing.T) {
+	cfg := &config.Config{
+		Settings:   config.Settings{TableName: "test", AddressFamily: config.FamilyINET},
+		Zones:      map[string]config.Zone{"fw": {Type: config.ZoneFirewall}, "net": {Type: config.ZoneIP}},
+		Interfaces: []config.Interface{{Zone: "net", Interface: "eth0"}},
+		Policy:     []config.Policy{{Source: "net", Dest: "all", Action: config.PolicyDrop, Log: "info"}},
+		PortGroups: make(map[string]config.PortGroup),
+	}
+	state, err := NewCompiler(cfg).Compile()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, r := range state.Rules["input"] {
+		for _, e := range r.Exprs {
+			if _, ok := e.(*expr.Limit); ok {
+				t.Errorf("%s: unexpected limit without log_limit", r.Tag)
+			}
+		}
+	}
+}
+
+func TestCompile_LogLimitSplitsAroundExtrasAndNAT(t *testing.T) {
+	cfg := listCfg(func(c *config.Config) {
+		c.Settings.LogLimit = "1/sec:10"
+		c.Zones["loc"] = config.Zone{Type: config.ZoneIP}
+		c.Interfaces = append(c.Interfaces, config.Interface{Zone: "loc", Interface: "eth1"})
+		c.Rules = []config.Rule{
+			{Action: config.RuleAccept, Source: "net:192.0.2.1", Dest: "fw", Proto: "tcp", DPort: config.PortSpec{"22"}, Log: "info",
+				Mark: "0x1", User: "0", Time: &config.TimeSpec{Start: "08:00", Stop: "17:00"}, RateLimit: "5/sec:20"},
+			{Action: config.RuleLog, Source: "net:192.0.2.2", Dest: "fw", Proto: "tcp", DPort: config.PortSpec{"23"}, Log: "info",
+				Mark: "0x2", RateLimit: "5/sec:20"},
+			{Action: config.RuleDNAT, Source: "net", Dest: "loc:198.51.100.10:80", Proto: "tcp", DPort: config.PortSpec{"8000"}, Log: "info"},
+			{Action: config.RuleAccept, Source: "net:192.0.2.3", Dest: "loc", Proto: "tcp", DPort: config.PortSpec{"443"}, Log: "info"},
+		}
+		c.SNAT = []config.SNATRule{{Action: config.SNATAddress, Source: "198.51.100.0/24", Dest: "eth0", Address: "203.0.113.7", Mark: "0x3", Log: "info"}}
+	})
+	state, err := NewCompiler(cfg).Compile()
+	if err != nil {
+		t.Fatalf("Compile() error: %v", err)
+	}
+	logLimit := &expr.Limit{Type: expr.LimitTypePkts, Rate: 1, Unit: expr.LimitTimeSecond, Burst: 10}
+	kinds := func(exprs []expr.Any) (logs, limits, marks, uids int) {
+		for _, e := range exprs {
+			switch v := e.(type) {
+			case *expr.Log:
+				logs++
+			case *expr.Limit:
+				limits++
+			case *expr.Meta:
+				if v.Key == expr.MetaKeyMARK && !v.SourceRegister {
+					marks++
+				}
+				if v.Key == expr.MetaKeySKUID {
+					uids++
+				}
+			}
+		}
+		return
+	}
+	check := func(name string, rs []ManagedRule, n, wantMarks, wantUIDs int) []expr.Any {
+		t.Helper()
+		if len(rs) != n {
+			t.Fatalf("%s: %d rules, want %d", name, len(rs), n)
+		}
+		l := rs[0].Exprs
+		if len(l) < 2 || !reflect.DeepEqual(l[len(l)-2], logLimit) {
+			t.Errorf("%s: log rule must end limit+log, got %#v", name, l)
+		}
+		if logs, limits, marks, uids := kinds(l); logs != 1 || limits != 1 || marks != wantMarks || uids != wantUIDs {
+			t.Errorf("%s: log rule logs=%d limits=%d marks=%d uids=%d, want 1/1/%d/%d", name, logs, limits, marks, uids, wantMarks, wantUIDs)
+		}
+		if n == 1 {
+			return nil
+		}
+		v := rs[1].Exprs
+		if logs, _, marks, uids := kinds(v); logs != 0 || marks != wantMarks || uids != wantUIDs {
+			t.Errorf("%s: action rule logs=%d marks=%d uids=%d, want 0/%d/%d", name, logs, marks, uids, wantMarks, wantUIDs)
+		}
+		return v
+	}
+
+	v := check("rule:0", taggedRules(state, "input", "rule:0"), 2, 1, 1)
+	if _, limits, _, _ := kinds(v); limits != 1 {
+		t.Errorf("rule:0: action rule must keep its ratelimit, got %d limits", limits)
+	}
+	if _, ok := v[len(v)-1].(*expr.Verdict); !ok {
+		t.Errorf("rule:0: action rule must end in a verdict, got %T", v[len(v)-1])
+	}
+	check("rule:1", taggedRules(state, "input", "rule:1"), 1, 1, 0)
+	if v := check("rule:2", taggedRules(state, "prerouting", "rule:2"), 2, 0, 0); v != nil {
+		if _, ok := v[len(v)-1].(*expr.NAT); !ok {
+			t.Errorf("rule:2: DNAT rule must end in nat, got %T", v[len(v)-1])
+		}
+	}
+	check("rule:3", taggedRules(state, "forward", "rule:3"), 2, 0, 0)
+	var snat []ManagedRule
+	for _, r := range state.Rules["postrouting"] {
+		if logs, _, _, _ := kinds(r.Exprs); logs > 0 || strings.HasPrefix(r.Tag, "snat:0") {
+			snat = append(snat, r)
+		}
+	}
+	if v := check("snat:0", snat, 2, 1, 0); v != nil {
+		if _, ok := v[len(v)-1].(*expr.NAT); !ok {
+			t.Errorf("snat:0: rule must end in nat, got %T", v[len(v)-1])
+		}
+	}
+}

internal/shorewall/convert.go

@@ -2,6 +2,7 @@ package shorewall
 
 import (
 	"fmt"
+	"log/slog"
 	"strconv"
 	"strings"
 
@@ -141,6 +142,17 @@ func convertConf(dir string, cfg *config.Config, params map[string]string, ipv6
 	} else {
 		cfg.Settings.LogLevel = "info"
 	}
+	if v := conf["LOGLIMIT"]; v != "" {
+		if strings.HasPrefix(v, "s:") || strings.HasPrefix(v, "d:") {
+			slog.Warn("shorewall: per-address LOGLIMIT is not supported, limiting each log site globally", "loglimit", v)
+			v = v[2:]
+		}
+		if name, rest, ok := strings.Cut(v, ":"); ok && !strings.Contains(name, "/") {
+			slog.Warn("shorewall: named LOGLIMIT is not supported, dropping the name", "loglimit", conf["LOGLIMIT"])
+			v = rest
+		}
+		cfg.Settings.LogLimit = v
+	}
 	if v, ok := conf["IP_FORWARDING"]; ok {
 		cfg.Settings.IPForwarding = v == "Yes" || v == "On" || v == "on" || v == "Keep"
 	}

internal/shorewall/convert_test.go

@@ -756,3 +756,23 @@ func TestConvert_Dispositions(t *testing.T) {
 		}
 	}
 }
+
+func TestConvert_LogLimit(t *testing.T) {
+	for in, want := range map[string]string{
+		`LOGLIMIT="s:1/sec:10"`:   "1/sec:10",
+		`LOGLIMIT=2/min`:          "2/min",
+		`LOGLIMIT=name:1/sec:5`:   "1/sec:5",
+		`LOGLIMIT=s:name:1/sec:5`: "1/sec:5",
+		`LOGLIMIT=`:               "",
+	} {
+		dir := minimalShorewallDir(t)
+		writeFile(t, dir, "shorewall.conf", "LOG_LEVEL=info\n"+in+"\n")
+		cfg, err := Convert(dir)
+		if err != nil {
+			t.Fatalf("%s: %v", in, err)
+		}
+		if cfg.Settings.LogLimit != want {
+			t.Errorf("%s: log_limit = %q, want %q", in, cfg.Settings.LogLimit, want)
+		}
+	}
+}

internal/tryapply/tryapply.go

@@ -25,15 +25,15 @@ const Unit = "tomswall-try-revert"
 var (
 	// Dir holds the lock and the pending snapshot.
 	Dir = "/var/lib/tomswall"
-	// run executes a systemd command; replaced in tests.
+	// Run executes a systemd command. Test hook; production code must not reassign.
-	run = func(name string, args ...string) error {
+	Run = func(name string, args ...string) error {
 		if out, err := exec.Command(name, args...).CombinedOutput(); err != nil {
 			return fmt.Errorf("%s %s: %w: %s", name, strings.Join(args, " "), err, strings.TrimSpace(string(out)))
 		}
 		return nil
 	}
-	// restore rolls the live table back to a snapshot; replaced in tests.
+	// Restore rolls the live table back to a snapshot. Test hook; production code must not reassign.
-	restore = func(s *nftables.Snapshot) error {
+	Restore = func(s *nftables.Snapshot) error {
 		engine, err := nftables.NewEngine(&config.Config{Settings: config.Settings{TableName: s.Table}})
 		if err != nil {
 			return err
@@ -94,23 +94,7 @@ func Arm(snap *nftables.Snapshot, pid int, delay time.Duration) (string, error)
 	if err != nil {
 		return "", err
 	}
-	f, err := os.CreateTemp(Dir, ".try-snapshot-*")
+	if err := WriteFile(snapshotPath(), b); err != nil {
-	if err != nil {
-		return "", err
-	}
-	defer os.Remove(f.Name())
-	if _, err := f.Write(b); err != nil {
-		f.Close()
-		return "", err
-	}
-	if err := f.Sync(); err != nil {
-		f.Close()
-		return "", err
-	}
-	if err := f.Close(); err != nil {
-		return "", err
-	}
-	if err := os.Rename(f.Name(), snapshotPath()); err != nil {
 		return "", err
 	}
 
@@ -119,13 +103,46 @@ func Arm(snap *nftables.Snapshot, pid int, delay time.Duration) (string, error)
 		return "", discardWith(err)
 	}
 	_ = disarm() // a leftover timer from an earlier try would block the unit name
-	if err := run("systemd-run", "--quiet", "--collect", "--unit", Unit,
+	if err := Run("systemd-run", "--quiet", "--collect", "--unit", Unit,
 		fmt.Sprintf("--on-active=%ds", int(delay.Round(time.Second).Seconds())), exe, "revert", "--id", id); err != nil {
 		return "", discardWith(fmt.Errorf("arming revert timer: %w", err))
 	}
 	return id, nil
 }
 
+// WriteFile durably replaces path with b: temp file, fsync, rename, fsync the directory.
+func WriteFile(path string, b []byte) error {
+	dir := filepath.Dir(path)
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return err
+	}
+	f, err := os.CreateTemp(dir, "."+filepath.Base(path)+"-*")
+	if err != nil {
+		return err
+	}
+	defer os.Remove(f.Name())
+	if _, err := f.Write(b); err != nil {
+		f.Close()
+		return err
+	}
+	if err := f.Sync(); err != nil {
+		f.Close()
+		return err
+	}
+	if err := f.Close(); err != nil {
+		return err
+	}
+	if err := os.Rename(f.Name(), path); err != nil {
+		return err
+	}
+	d, err := os.Open(dir)
+	if err != nil {
+		return err
+	}
+	defer d.Close()
+	return d.Sync()
+}
+
 // Discard drops the pending snapshot and timer without restoring. The caller must hold the lock.
 func Discard() error {
 	_ = disarm()
@@ -143,7 +160,7 @@ func discardWith(err error) error {
 }
 
 func disarm() error {
-	return run("systemctl", "stop", Unit+".timer")
+	return Run("systemctl", "stop", Unit+".timer")
 }
 
 // Confirm keeps the tried ruleset. ok is false when no try was pending, i.e.
@@ -193,7 +210,7 @@ func Abort() error {
 }
 
 func restorePending(p *pending) error {
-	if err := restore(p.Snapshot); err != nil {
+	if err := Restore(p.Snapshot); err != nil {
 		return fmt.Errorf("restoring snapshot: %w", err)
 	}
 	return Discard()

internal/tryapply/tryapply_test.go

@@ -15,12 +15,12 @@ func setup(t *testing.T) *[]string {
 	t.Helper()
 	Dir = t.TempDir()
 	var cmds []string
-	orig := run
+	orig := Run
-	run = func(name string, args ...string) error {
+	Run = func(name string, args ...string) error {
 		cmds = append(cmds, name+" "+strings.Join(args, " "))
 		return nil
 	}
-	t.Cleanup(func() { run = orig })
+	t.Cleanup(func() { Run = orig })
 	return &cmds
 }
 
@@ -91,7 +91,7 @@ func TestAcquireRefusesWhilePending(t *testing.T) {
 
 func TestArmFailureDiscardsSnapshot(t *testing.T) {
 	setup(t)
-	run = func(name string, args ...string) error {
+	Run = func(name string, args ...string) error {
 		if name == "systemd-run" {
 			return errors.New("no systemd")
 		}
@@ -145,12 +145,12 @@ func TestConfirmAfterRevertFails(t *testing.T) {
 func stubRestore(t *testing.T, err error) *[]*nftables.Snapshot {
 	t.Helper()
 	var got []*nftables.Snapshot
-	orig := restore
+	orig := Restore
-	restore = func(s *nftables.Snapshot) error {
+	Restore = func(s *nftables.Snapshot) error {
 		got = append(got, s)
 		return err
 	}
-	t.Cleanup(func() { restore = orig })
+	t.Cleanup(func() { Restore = orig })
 	return &got
 }
 

packaging/nfpm.yaml

@@ -47,6 +47,17 @@ contents:
     file_info:
       mode: 0640
 
+  # systemd unit + environment file for applying a local config at boot.
+  - src: packaging/tomswall.service
+    dst: /usr/lib/systemd/system/tomswall.service
+    file_info:
+      mode: 0644
+  - src: packaging/tomswall.env
+    dst: /etc/tomswall/tomswall.env
+    type: config|noreplace
+    file_info:
+      mode: 0644
+
   # Shell completions (generated by scripts/build-rpm.sh before packaging).
   - src: dist/completions/tomswall.bash
     dst: /usr/share/bash-completion/completions/tomswall

packaging/tomswall-agent.service

@@ -3,6 +3,7 @@ Description=tomswall control-plane agent (pull and apply firewall config)
 Documentation=https://git.unkin.net/unkin/tomswall
 After=network-online.target
 Wants=network-online.target
+Conflicts=tomswall.service
 
 [Service]
 Type=simple

packaging/tomswall.env

@@ -0,0 +1,3 @@
+# Config applied by tomswall.service: a tomswall YAML file or a shorewall directory.
+TOMSWALL_CONFIG=/etc/tomswall/tomswall.yaml
+#TOMSWALL_CONFIG=/etc/shorewall

packaging/tomswall.service

@@ -0,0 +1,25 @@
+[Unit]
+Description=tomswall firewall (apply local config at boot)
+Documentation=https://git.unkin.net/unkin/tomswall
+DefaultDependencies=no
+Wants=network-pre.target
+Before=network-pre.target shutdown.target
+After=local-fs.target systemd-sysctl.service
+Conflicts=shutdown.target tomswall-agent.service
+StartLimitIntervalSec=60
+StartLimitBurst=5
+
+[Service]
+Type=oneshot
+RemainAfterExit=yes
+Environment=TOMSWALL_CONFIG=/etc/tomswall/tomswall.yaml
+EnvironmentFile=-/etc/tomswall/tomswall.env
+ExecStart=/usr/sbin/tomswall apply -c ${TOMSWALL_CONFIG}
+ExecReload=/usr/sbin/tomswall apply -c ${TOMSWALL_CONFIG}
+# Fails open: after StartLimitBurst failures within StartLimitIntervalSec, boot continues without the ruleset.
+Restart=on-failure
+RestartSec=5
+# No ExecStop: stopping the unit leaves the ruleset in place (flush would open the firewall).
+
+[Install]
+WantedBy=sysinit.target

tomswall.example.yaml

@@ -6,6 +6,8 @@ settings:
   address_family: inet
   ip_forwarding: true
   log_level: info
+  # rate limit for every log site (shorewall LOGLIMIT, global form): rate/{sec|min|hour|day}[:burst]; unset logs every hit
+  log_limit: 1/sec:10
   table_name: tomswall
   implicit_continue: false
   # ct state invalid/untracked verdict: accept, drop, reject, continue (pass to rules)