Redo the banning logic, ban.Service

CODE_OF_CONDUCT.md

@@ -60,7 +60,7 @@ representative at an online or offline event.
 
 Instances of abusive, harassing, or otherwise unacceptable behavior may be
 reported to the community leaders responsible for enforcement via Discord/Matrix (binwiederhier),
-or email (ntfy@heckel.io). All complaints will be reviewed and investigated promptly 
+or email (contact@mail.ntfy.sh). All complaints will be reviewed and investigated promptly 
 and fairly.
 
 All community leaders are obligated to respect the privacy and security of the

ban/service.go

@@ -0,0 +1,185 @@
+// Package ban implements the abuse ban-feed: it tracks per-prefix weighted "strikes" from rejected
+// requests and appends breaching prefixes to a ban file that fail2ban tails. Keying by prefix (not
+// by visitor) makes the accounting match the unit fail2ban bans, even for shared account visitors.
+package ban
+
+import (
+	"fmt"
+	"net/netip"
+	"os"
+	"sync"
+	"time"
+
+	"golang.org/x/time/rate"
+	"heckel.io/ntfy/v2/log"
+)
+
+const (
+	tag           = "ban"
+	pruneInterval = 10 * time.Minute
+	writeInterval = 3 * time.Second
+)
+
+// Config is the Service's config, kept separate from server.Config to avoid an import cycle.
+type Config struct {
+	File           string        // Ban file that fail2ban tails (must be non-empty; the caller decides whether the feature is enabled)
+	Window         time.Duration // Rolling window over which weighted strikes are counted
+	Threshold      int           // Weighted strikes per Window before a prefix is banned
+	Weights        Weights       // Code matcher -> strike weight (0 = exempt)
+	PrefixBitsIPv4 int           // Mask width for the ban unit, e.g. 32 (matches rate-limiting granularity)
+	PrefixBitsIPv6 int           // Mask width for the ban unit, e.g. 64
+}
+
+// tracker is the per-prefix strike state: a weighted breach detector plus timestamps for pruning and throttling.
+type tracker struct {
+	limiter *rate.Limiter
+	seen    time.Time // Last strike, for pruning idle prefixes
+	emitted time.Time // Last ban-line write for this prefix, throttles re-emits to once per Window
+}
+
+// Service owns the ban-feed: per-prefix strike accounting, buffered file writes, and idle-prefix
+// pruning. The caller owns the enable/disable decision -- only construct a Service when the feature
+// is on (see server.New, which builds one only when a ban file is configured).
+type Service struct {
+	conf      *Config
+	mu        sync.Mutex // Guards trackers and pending
+	trackers  map[netip.Prefix]*tracker
+	pending   []string      // Formatted ban lines buffered by Record, flushed to the ban file by runWriteLoop
+	writeDone chan struct{} // Closed when runWriteLoop exits after its final flush
+	closeChan chan struct{}
+	closeOnce sync.Once
+}
+
+// NewService builds a Service and starts its background loops. The caller must only call it when the
+// feature is enabled (conf non-nil, File non-empty); the Service does not model a disabled state.
+func NewService(conf *Config) *Service {
+	s := &Service{
+		conf:      conf,
+		trackers:  make(map[netip.Prefix]*tracker),
+		closeChan: make(chan struct{}),
+		writeDone: make(chan struct{}),
+	}
+	go s.runPruneLoop()
+	go s.runWriteLoop()
+	return s
+}
+
+// Record counts one rejection against the IP's prefix bucket and, on breach, buffers a ban line
+// (throttled to once per Window per prefix). No-ops for a non-4xx/5xx status or a zero-weight code.
+func (s *Service) Record(ip netip.Addr, httpCode, errorCode int) {
+	if httpCode < 400 {
+		return
+	}
+	weight := s.conf.Weights.WeightFor(errorCode)
+	if weight == 0 {
+		return // Weight 0: exempt, no strike
+	}
+	prefix := s.prefix(ip)
+	now := time.Now()
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	t := s.trackers[prefix]
+	if t == nil {
+		t = &tracker{limiter: rate.NewLimiter(rate.Limit(float64(s.conf.Threshold)/s.conf.Window.Seconds()), s.conf.Threshold)}
+		s.trackers[prefix] = t
+	}
+	t.seen = now
+	if t.limiter.AllowN(now, weight) {
+		return // Within the strike budget, no breach
+	}
+	if !t.emitted.IsZero() && now.Sub(t.emitted) < s.conf.Window {
+		return // Already emitted this prefix within the window (one ban line per prefix per window)
+	}
+	t.emitted = now
+	s.pending = append(s.pending, formatBanLine(now, ip, prefix, httpCode, errorCode))
+}
+
+// prefix masks ip to the ban unit (PrefixBitsIPv4/IPv6) -- what fail2ban bans, e.g. a whole /64.
+func (s *Service) prefix(ip netip.Addr) netip.Prefix {
+	if ip.Is4() {
+		return netip.PrefixFrom(ip, s.conf.PrefixBitsIPv4).Masked()
+	}
+	return netip.PrefixFrom(ip, s.conf.PrefixBitsIPv6).Masked()
+}
+
+// formatBanLine builds the "<RFC3339-UTC> <ip> <prefix> <http> <ntfy>" line the fail2ban filter
+// parses. The timestamp is captured at breach time, not flush time.
+func formatBanLine(t time.Time, ip netip.Addr, prefix netip.Prefix, httpCode, errorCode int) string {
+	return fmt.Sprintf("%s %s %s %d %d\n", t.UTC().Format(time.RFC3339), ip.String(), prefix.String(), httpCode, errorCode)
+}
+
+// runPruneLoop prunes idle prefixes until Close.
+func (s *Service) runPruneLoop() {
+	ticker := time.NewTicker(pruneInterval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ticker.C:
+			s.prune()
+		case <-s.closeChan:
+			return
+		}
+	}
+}
+
+// runWriteLoop flushes buffered ban lines every writeInterval, plus a final flush on Close.
+func (s *Service) runWriteLoop() {
+	defer close(s.writeDone)
+	ticker := time.NewTicker(writeInterval)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-ticker.C:
+			s.flush()
+		case <-s.closeChan:
+			s.flush()
+			return
+		}
+	}
+}
+
+// flush appends the buffered lines to the file in one open/write. Best-effort: a batch is dropped on
+// error. Concurrent calls are safe -- pending is drained under mu, so only one flush writes a batch.
+func (s *Service) flush() {
+	s.mu.Lock()
+	lines := s.pending
+	s.pending = nil
+	s.mu.Unlock()
+	if len(lines) == 0 {
+		return
+	}
+	f, err := os.OpenFile(s.conf.File, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
+	if err != nil {
+		log.Tag(tag).Err(err).Warn("Cannot open ban file %s, dropped %d ban(s)", s.conf.File, len(lines))
+		return
+	}
+	defer f.Close()
+	for i, line := range lines {
+		if _, err := f.WriteString(line); err != nil {
+			log.Tag(tag).Err(err).Warn("Cannot write to ban file %s, dropped %d ban(s)", s.conf.File, len(lines)-i)
+			return
+		}
+	}
+}
+
+// prune drops prefixes idle for a full Window -- their bucket has refilled, so forgetting them is a
+// no-op that bounds memory under a flood of distinct IPs.
+func (s *Service) prune() {
+	now := time.Now()
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for prefix, t := range s.trackers {
+		if now.Sub(t.seen) >= s.conf.Window {
+			delete(s.trackers, prefix)
+		}
+	}
+}
+
+// Close stops the loops and blocks until the final flush completes. Idempotent.
+func (s *Service) Close() {
+	s.closeOnce.Do(func() {
+		close(s.closeChan)
+		<-s.writeDone
+	})
+}

ban/service_test.go

@@ -0,0 +1,253 @@
+package ban
+
+import (
+	"net/netip"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+var testIP = netip.MustParseAddr("1.2.3.4")
+
+// newTestService creates a Service wired for testing, with the given ban file, weighted-bucket
+// threshold, and weights, plus a 1-minute window (so the emit throttle only fires once per test).
+func newTestService(t *testing.T, banFile string, threshold int, weights map[string]int) *Service {
+	t.Helper()
+	s := NewService(&Config{
+		File:           banFile,
+		Window:         time.Minute,
+		Threshold:      threshold,
+		Weights:        weights,
+		PrefixBitsIPv4: 32,
+		PrefixBitsIPv6: 64,
+	})
+	t.Cleanup(s.Close)
+	return s
+}
+
+// flushAndRead forces a synchronous flush of the buffered bans (writes are otherwise async, on the
+// runWriteLoop ticker) and returns the ban file's lines.
+func flushAndRead(t *testing.T, s *Service, path string) []string {
+	t.Helper()
+	s.flush()
+	data, err := os.ReadFile(path)
+	require.NoError(t, err)
+	return strings.Split(strings.TrimRight(string(data), "\n"), "\n")
+}
+
+func TestService_Record_Weight2BansAtHalfThreshold(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	// Threshold 10, code weight 2 -> the budget covers exactly 5 hits, so the 6th breaches.
+	s := newTestService(t, banFile, 10, map[string]int{"*": 2})
+	for i := 0; i < 5; i++ {
+		s.Record(testIP, 400, 40001)
+	}
+	s.flush()
+	require.NoFileExists(t, banFile) // 5 hits * weight 2 = 10 == budget, exactly at the limit, not over
+	s.Record(testIP, 400, 40001)     // 6th hit cannot be covered -> breach
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 400 40001")) // <ip> <prefix> <http> <ntfy-code>
+}
+
+func TestService_Record_Weight10BansFast(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	// Threshold 10, code weight 10 -> a single hit drains the whole budget, so the 2nd breaches.
+	s := newTestService(t, banFile, 10, map[string]int{"42909": 10, "*": 1})
+	s.Record(testIP, 429, 42909)
+	s.flush()
+	require.NoFileExists(t, banFile)
+	s.Record(testIP, 429, 42909) // 2nd hit cannot be covered -> breach
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 429 42909"))
+}
+
+func TestService_Record_Weight0NeverBans(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	// The legit-quota code is exempt (weight 0), so no number of hits ever bans.
+	s := newTestService(t, banFile, 10, map[string]int{"42908": 0, "*": 1})
+	for i := 0; i < 100; i++ {
+		s.Record(testIP, 429, 42908)
+	}
+	s.flush()
+	require.NoFileExists(t, banFile)
+}
+
+func TestService_Record_SingleBucketNoRelaxation(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	// One shared bucket per prefix: different codes draw down the SAME budget, so mixing them creates
+	// no extra headroom (unlike per-code buckets, which would relax the effective limit for a mixed
+	// offender).
+	s := newTestService(t, banFile, 10, map[string]int{"403*": 2, "*": 1})
+	s.Record(testIP, 403, 40301) // weight 2 -> 8 left
+	s.Record(testIP, 403, 40301) // weight 2 -> 6 left
+	s.Record(testIP, 403, 40301) // weight 2 -> 4 left
+	s.flush()
+	require.NoFileExists(t, banFile)
+	for i := 0; i < 4; i++ {
+		s.Record(testIP, 400, 40001) // weight 1 each -> drains the remaining 4 -> 0 left
+	}
+	s.flush()
+	require.NoFileExists(t, banFile) // 6 + 4 = 10 == budget exactly, still not over
+	s.Record(testIP, 400, 40001)     // one more cannot be covered -> breach
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+}
+
+func TestService_Record_ExactLineFormat(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 1, map[string]int{"*": 1})
+	before := time.Now().UTC().Truncate(time.Second)
+	for i := 0; i < 3; i++ {
+		s.Record(testIP, 429, 42901)
+	}
+	after := time.Now().UTC()
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+	parts := strings.Split(lines[0], " ")
+	require.Len(t, parts, 5) // "<timestamp> <ip> <prefix> <http-code> <ntfy-code>"
+	ts, err := time.Parse(time.RFC3339, parts[0])
+	require.NoError(t, err)
+	require.Equal(t, time.UTC, ts.Location())
+	require.False(t, ts.Before(before))
+	require.False(t, ts.After(after.Add(time.Second)))
+	require.Equal(t, "1.2.3.4", parts[1])    // full IP
+	require.Equal(t, "1.2.3.4/32", parts[2]) // masked to the default IPv4 prefix (/32)
+	require.Equal(t, "429", parts[3])        // HTTP status
+	require.Equal(t, "42901", parts[4])      // ntfy code
+}
+
+func TestService_Record_IPv6MaskedToPrefix(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 1, map[string]int{"*": 1})
+	ip := netip.MustParseAddr("2001:db8::abcd")
+	for i := 0; i < 3; i++ {
+		s.Record(ip, 429, 42901)
+	}
+	parts := strings.Split(flushAndRead(t, s, banFile)[0], " ")
+	require.Len(t, parts, 5)
+	require.Equal(t, "2001:db8::abcd", parts[1]) // full IPv6 address
+	require.Equal(t, "2001:db8::/64", parts[2])  // masked to the default IPv6 prefix (/64)
+}
+
+func TestService_Record_PerPrefixIsolation(t *testing.T) {
+	// Each source prefix gets its own bucket: one IP hammering to a breach must not push a different,
+	// quiet IP over the edge. This is the whole point of keying by prefix instead of by visitor.
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 2, map[string]int{"*": 1})
+	noisy := netip.MustParseAddr("1.1.1.1")
+	quiet := netip.MustParseAddr("2.2.2.2")
+	for i := 0; i < 5; i++ {
+		s.Record(noisy, 429, 42901) // breaches its own bucket
+	}
+	s.Record(quiet, 429, 42901) // single hit, well under threshold
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1) // only the noisy prefix is written
+	require.True(t, strings.HasSuffix(lines[0], " 1.1.1.1 1.1.1.1/32 429 42901"))
+}
+
+func TestService_Record_OncePerWindowThrottle(t *testing.T) {
+	// Once a prefix has been written, further breaches within the window must not re-append it, so a
+	// persistent offender produces exactly one line per window (mirrors the old per-visitor banEmit).
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 1, map[string]int{"*": 1})
+	for i := 0; i < 50; i++ {
+		s.Record(testIP, 429, 42901)
+	}
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+}
+
+func TestService_Record_BansPassedIP(t *testing.T) {
+	// The Service bans the exact IP passed to Record -- the caller passes the offending request's IP,
+	// which for an account-keyed visitor is not the visitor's stored IP.
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 1, map[string]int{"*": 1})
+	offender := netip.MustParseAddr("5.6.7.8")
+	for i := 0; i < 3; i++ {
+		s.Record(offender, 429, 42901)
+	}
+	parts := strings.Split(flushAndRead(t, s, banFile)[0], " ")
+	require.Equal(t, "5.6.7.8", parts[1])    // the IP passed to Record
+	require.Equal(t, "5.6.7.8/32", parts[2]) // its prefix
+}
+
+func TestService_Record_Ignores2xx3xx(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 3, map[string]int{"*": 1})
+	// Success and redirects must never count toward a ban, even over the threshold -- otherwise a
+	// legit high-volume publisher (lots of 200s) would get banned.
+	for i := 0; i < 20; i++ {
+		s.Record(testIP, 200, 20000)
+		s.Record(testIP, 302, 30000)
+	}
+	s.flush()
+	require.NoFileExists(t, banFile)
+	// A 4xx over the same budget still gets written.
+	for i := 0; i < 5; i++ {
+		s.Record(testIP, 400, 40001)
+	}
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 400 40001"))
+}
+
+func TestService_Record_BuffersUntilFlush(t *testing.T) {
+	// Writes are async: a breach buffers the ban line rather than writing it synchronously on the
+	// request path. The line only reaches the file when runWriteLoop (or an explicit flush) runs.
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 1, map[string]int{"*": 1})
+	for i := 0; i < 3; i++ {
+		s.Record(testIP, 429, 42901)
+	}
+	require.NoFileExists(t, banFile) // not written synchronously
+	s.mu.Lock()
+	require.Len(t, s.pending, 1) // one line buffered (throttled to once per window)
+	s.mu.Unlock()
+	lines := flushAndRead(t, s, banFile)
+	require.Len(t, lines, 1)
+	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 429 42901"))
+}
+
+func TestService_Close_FlushesPending(t *testing.T) {
+	// Close must flush buffered bans so nothing is lost on shutdown, and must block until it has.
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := NewService(&Config{File: banFile, Window: time.Minute, Threshold: 1, Weights: Weights{"*": 1}, PrefixBitsIPv4: 32, PrefixBitsIPv6: 64})
+	for i := 0; i < 3; i++ {
+		s.Record(testIP, 429, 42901)
+	}
+	require.NoFileExists(t, banFile) // still buffered
+	s.Close()                        // blocks until the final flush completes
+	data, err := os.ReadFile(banFile)
+	require.NoError(t, err)
+	require.Len(t, strings.Split(strings.TrimRight(string(data), "\n"), "\n"), 1)
+}
+
+func TestService_Prune_DropsIdlePrefixes(t *testing.T) {
+	// A prefix idle for a full window has a refilled bucket, so prune drops it to bound memory. An
+	// active prefix (seen within the window) is kept.
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	s := newTestService(t, banFile, 10, map[string]int{"*": 1})
+	idle := netip.MustParseAddr("9.9.9.9")
+	s.Record(idle, 400, 40001)
+	s.Record(testIP, 400, 40001)
+	require.Len(t, s.trackers, 2)
+
+	// Backdate the idle prefix past the window, then prune.
+	idlePrefix := s.prefix(idle)
+	s.mu.Lock()
+	s.trackers[idlePrefix].seen = time.Now().Add(-2 * time.Minute)
+	s.mu.Unlock()
+	s.prune()
+
+	require.Len(t, s.trackers, 1)
+	_, ok := s.trackers[idlePrefix]
+	require.False(t, ok) // idle prefix dropped
+	_, ok = s.trackers[s.prefix(testIP)]
+	require.True(t, ok) // active prefix kept
+}

ban/weights.go

@@ -0,0 +1,83 @@
+package ban
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+)
+
+// Weights maps a matcher key to a strike weight for the abuse ban-feed (see ParseWeights, WeightFor).
+type Weights map[string]int
+
+// ParseWeights normalizes a list like ["42909:10","403:2","*:1"] into a Weights map. A key is an
+// exact ntfy code, a family ("429*"), a bare HTTP status ("403" -> "403*"), or "*"; weights are ints
+// >= 0 (0 = exempt). Malformed entries are rejected so misconfiguration fails at startup.
+func ParseWeights(entries []string) (Weights, error) {
+	out := make(Weights, len(entries))
+	for _, entry := range entries {
+		key, weightStr, ok := strings.Cut(entry, ":")
+		if !ok {
+			return nil, fmt.Errorf("invalid ban-weight %q, want KEY:WEIGHT", entry)
+		}
+		weight, err := strconv.Atoi(strings.TrimSpace(weightStr))
+		if err != nil || weight < 0 {
+			return nil, fmt.Errorf("invalid ban-weight value in %q, want a non-negative integer", entry)
+		}
+		key = strings.TrimSpace(key)
+		if !validWeightKey(key) {
+			return nil, fmt.Errorf("invalid ban-weight key in %q, want %q, an ntfy code, an HTTP status, or a PREFIX*", entry, "*")
+		}
+		// A bare 3-digit HTTP status is shorthand for the whole family (e.g. "403" -> "403*").
+		if len(key) == 3 && isAllDigits(key) {
+			key += "*"
+		}
+		out[key] = weight
+	}
+	return out, nil
+}
+
+// WeightFor returns the strike weight for an ntfy error code, longest-match-wins (exact > family > "*").
+// If nothing matches (no "*" catch-all) it returns the implied default 1, so a forgotten "*" still
+// bans; use "*:0" to exempt everything not explicitly weighted.
+func (w Weights) WeightFor(errorCode int) int {
+	code := strconv.Itoa(errorCode)
+	weight, bestLen, matched := 0, -1, false
+	for key, wt := range w {
+		matchLen := -1
+		switch {
+		case key == "*":
+			matchLen = 0
+		case strings.HasSuffix(key, "*"):
+			if prefix := strings.TrimSuffix(key, "*"); strings.HasPrefix(code, prefix) {
+				matchLen = len(prefix)
+			}
+		case key == code:
+			matchLen = len(code)
+		}
+		if matchLen > bestLen {
+			weight, bestLen, matched = wt, matchLen, true
+		}
+	}
+	if !matched {
+		return 1
+	}
+	return weight
+}
+
+// validWeightKey reports whether key is a legal matcher: "*", an all-digits code, or DIGITS*.
+func validWeightKey(key string) bool {
+	if key == "*" {
+		return true
+	}
+	digits := strings.TrimSuffix(key, "*")
+	return digits != "" && isAllDigits(digits)
+}
+
+func isAllDigits(s string) bool {
+	for _, r := range s {
+		if r < '0' || r > '9' {
+			return false
+		}
+	}
+	return s != ""
+}

ban/weights_test.go

@@ -0,0 +1,73 @@
+package ban
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestParseWeights(t *testing.T) {
+	// Exact codes, a bare 3-digit HTTP status (normalized to a family), an exempt code, and "*".
+	weights, err := ParseWeights([]string{"42909:10", "403:2", "42908:0", "*:1"})
+	require.NoError(t, err)
+	require.Equal(t, Weights{"42909": 10, "403*": 2, "42908": 0, "*": 1}, weights)
+
+	// A bare 3-digit HTTP status normalizes to its family.
+	weights, err = ParseWeights([]string{"429:5"})
+	require.NoError(t, err)
+	require.Equal(t, Weights{"429*": 5}, weights)
+
+	// An explicit family key stays as-is.
+	weights, err = ParseWeights([]string{"429*:5"})
+	require.NoError(t, err)
+	require.Equal(t, Weights{"429*": 5}, weights)
+
+	// Weight 0 is valid and means exempt.
+	weights, err = ParseWeights([]string{"42908:0"})
+	require.NoError(t, err)
+	require.Equal(t, Weights{"42908": 0}, weights)
+
+	_, err = ParseWeights([]string{"401"}) // Missing weight
+	require.Error(t, err)
+	_, err = ParseWeights([]string{"401:-1"}) // Negative weight
+	require.Error(t, err)
+	_, err = ParseWeights([]string{"401:abc"}) // Non-integer weight
+	require.Error(t, err)
+	_, err = ParseWeights([]string{"abc:10"}) // Non-numeric key
+	require.Error(t, err)
+	_, err = ParseWeights([]string{"4*3:10"}) // Star not at the end
+	require.Error(t, err)
+}
+
+func TestWeights_WeightFor(t *testing.T) {
+	weights, err := ParseWeights([]string{"42908:0", "42903:0", "42905:0", "42910:0", "42909:10", "429*:1", "403*:2", "4*:1", "5*:1"})
+	require.NoError(t, err)
+	// Longest-match-wins: exact 5-digit beats "429*" beats "4*" beats "*".
+	require.Equal(t, 0, weights.WeightFor(42908))
+	require.Equal(t, 10, weights.WeightFor(42909))
+	require.Equal(t, 1, weights.WeightFor(42901))
+	require.Equal(t, 2, weights.WeightFor(40311))
+	require.Equal(t, 1, weights.WeightFor(40011))
+	require.Equal(t, 1, weights.WeightFor(50312))
+	// No rule matches (this config has 4*/5* but no "*"), so the implied default weight 1 applies.
+	require.Equal(t, 1, weights.WeightFor(30012))
+}
+
+func TestWeights_WeightFor_NoStarRuleImpliesWeight1(t *testing.T) {
+	// With no "*" rule, a code that matches nothing defaults to weight 1 (can be banned), so the
+	// feature can't be silently turned into a no-op by forgetting "*". Explicit codes still win.
+	weights, err := ParseWeights([]string{"42908:0", "42909:10"})
+	require.NoError(t, err)
+	require.Equal(t, 0, weights.WeightFor(42908))  // explicitly exempt
+	require.Equal(t, 10, weights.WeightFor(42909)) // explicit
+	require.Equal(t, 1, weights.WeightFor(42901))  // unmatched -> implied 1
+	require.Equal(t, 1, weights.WeightFor(40001))  // unmatched -> implied 1
+}
+
+func TestWeights_WeightFor_ExplicitStarZeroExemptsAll(t *testing.T) {
+	// An explicit "*:0" is the opt-out: exempt everything not otherwise weighted.
+	weights, err := ParseWeights([]string{"42909:10", "*:0"})
+	require.NoError(t, err)
+	require.Equal(t, 10, weights.WeightFor(42909)) // explicit
+	require.Equal(t, 0, weights.WeightFor(42901))  // *:0 -> exempt everything else
+}

cmd/serve.go

@@ -18,6 +18,7 @@ import (
 
 	"github.com/urfave/cli/v2"
 	"github.com/urfave/cli/v2/altsrc"
+	"heckel.io/ntfy/v2/ban"
 	"heckel.io/ntfy/v2/log"
 	"heckel.io/ntfy/v2/payments"
 	"heckel.io/ntfy/v2/server"
@@ -101,7 +102,7 @@ var flagsServe = append(
 	altsrc.NewIntFlag(&cli.IntFlag{Name: "visitor-prefix-bits-ipv6", Aliases: []string{"visitor_prefix_bits_ipv6"}, EnvVars: []string{"NTFY_VISITOR_PREFIX_BITS_IPV6"}, Value: server.DefaultVisitorPrefixBitsIPv6, Usage: "number of bits of the IPv6 address to use for rate limiting (default: 64, /64 subnet)"}),
 	altsrc.NewStringFlag(&cli.StringFlag{Name: "ban-file", Aliases: []string{"ban_file"}, EnvVars: []string{"NTFY_BAN_FILE"}, Value: "", Usage: "if set, append IPs of abusive visitors to this file for fail2ban to tail (empty disables)"}),
 	altsrc.NewStringFlag(&cli.StringFlag{Name: "ban-window", Aliases: []string{"ban_window"}, EnvVars: []string{"NTFY_BAN_WINDOW"}, Value: util.FormatDuration(server.DefaultBanWindow), Usage: "rolling window over which weighted strikes are counted for the ban file"}),
-	altsrc.NewIntFlag(&cli.IntFlag{Name: "ban-threshold", Aliases: []string{"ban_threshold"}, EnvVars: []string{"NTFY_BAN_THRESHOLD"}, Value: server.DefaultBanThreshold, Usage: "weighted strikes per window before a visitor is banned"}),
+	altsrc.NewIntFlag(&cli.IntFlag{Name: "ban-threshold", Aliases: []string{"ban_threshold"}, EnvVars: []string{"NTFY_BAN_THRESHOLD"}, Value: server.DefaultBanThreshold, Usage: "weighted strikes per window before an offender is banned"}),
 	altsrc.NewStringSliceFlag(&cli.StringSliceFlag{Name: "ban-weights", Aliases: []string{"ban_weights"}, EnvVars: []string{"NTFY_BAN_WEIGHTS"}, Value: cli.NewStringSlice(server.DefaultBanWeights...), Usage: "per-code strike weights as KEY:WEIGHT, where KEY is an ntfy code, an HTTP status, a PREFIX*, or '*' (weight 0 exempts)"}),
 	altsrc.NewBoolFlag(&cli.BoolFlag{Name: "behind-proxy", Aliases: []string{"behind_proxy", "P"}, EnvVars: []string{"NTFY_BEHIND_PROXY"}, Value: false, Usage: "if set, use forwarded header (e.g. X-Forwarded-For, X-Client-IP) to determine visitor IP address (for rate limiting)"}),
 	altsrc.NewStringFlag(&cli.StringFlag{Name: "proxy-forwarded-header", Aliases: []string{"proxy_forwarded_header"}, EnvVars: []string{"NTFY_PROXY_FORWARDED_HEADER"}, Value: "X-Forwarded-For", Usage: "use specified header to determine visitor IP address (for rate limiting)"}),
@@ -285,7 +286,7 @@ func execServe(c *cli.Context) error {
 	}
 
 	// Parse abuse ban-feed weights ("KEY:WEIGHT" list, "*" fallback)
-	banWeights, err := server.ParseBanWeights(banWeightsRaw)
+	banWeights, err := ban.ParseWeights(banWeightsRaw)
 	if err != nil {
 		return err
 	}

docs/config.md

@@ -2129,6 +2129,74 @@ chain.
 The official ntfy.sh server uses fail2ban to ban IPs. Check out ntfy.sh's [Ansible fail2ban role](https://github.com/binwiederhier/ntfy-ansible/tree/main/roles/fail2ban) for details. Ban actors are banned for 1 hour initially, and up to
 4 hours at a time for repeated offenses. IPv4 addresses are banned individually, while IPv6 addresses are banned by their `/56` prefix.
 
+#### Ban-feed
+In addition to the fail2ban setup above, ntfy can detect abusive visitors itself and write their IP
+addresses to a file for fail2ban to ban from. ntfy keeps a per-prefix weighted "strike" budget, and
+each rejected request costs strikes based on its response code -- the ntfy error code, or its HTTP
+status (see `ban-weights`) -- so different kinds of rejection can be weighted differently or exempted
+entirely. When a prefix exceeds the budget, ntfy appends the offending IP address to `ban-file`. Since
+every line is already a confirmed offender, the fail2ban jail can ban on first sight (`maxretry = 1`).
+
+- `ban-file` is the file offenders are appended to. If it is not set, the ban-feed is disabled. Its
+  parent directory must exist and be writable by ntfy. Be sure to rotate it (e.g. with logrotate and
+  `copytruncate`) so it does not grow unbounded.
+- `ban-window` is the rolling window over which weighted strikes are counted, per IP prefix.
+- `ban-threshold` is the number of weighted strikes per `ban-window` before a prefix is written to
+  `ban-file`. Each prefix has one shared budget, so it cannot be gamed by mixing error codes.
+- `ban-weights` assigns a strike weight per matcher key, formatted as `KEY:WEIGHT`. A key is an exact
+  ntfy error code (`42909`), a code family (`429*`, `403*`, `4*`), a bare HTTP status (`403`, short for
+  `403*`), or `*`. The longest matching key wins. A weight of `0` exempts a code, which is useful to
+  spare legit plan/quota `429`s from a `*` catch-all. Heavier weights ban faster. If you do not
+  include a `*` rule, any code that matches nothing defaults to weight `1` (i.e. it can be banned);
+  set `*:0` to exempt everything that is not explicitly weighted.
+
+Only rejections (4xx/5xx) count towards a ban; successful requests never do. Because the budget
+refills over `ban-window`, the trigger is a sustained rate: a prefix is only written out once it
+exceeds `ban-threshold / ban-window` rejections per second (with the defaults, `100 / 10m` = ~0.17/s).
+
+Each line in `ban-file` has the format `<RFC3339-timestamp> <ip> <prefix> <http-code> <ntfy-code>`, for example:
+
+```
+2026-01-15T20:56:32Z 1.2.3.4 1.2.3.4/32 429 42901
+2026-01-15T20:56:32Z 2001:db8::abcd 2001:db8::/64 429 42909
+```
+
+`<prefix>` is `<ip>` masked to the rate-limiting prefix (`visitor-prefix-bits-ipv4`/`-ipv6`) -- the
+same unit ntfy rate-limits by. Have the fail2ban filter capture the bare `<ip>` (the action then
+applies the prefix):
+
+=== "server.yml"
+    ```yaml
+    ban-file: "/var/log/ntfy/ban.log"
+    ban-window: "10m"
+    ban-threshold: 100
+    ban-weights:
+      - "42909:10"   # too many auth failures -> brute force, ban fast
+      - "42908:0"    # daily message quota reached -> legit, never counts
+      # everything else 4xx/5xx defaults to weight 1
+    ```
+
+=== "/etc/fail2ban/filter.d/ntfy-ban.conf"
+    ```
+    [Definition]
+    failregex = ^\S+ <HOST> \S+ \d+ \d+$
+    datepattern = ^%%Y-%%m-%%dT%%H:%%M:%%S
+    ignoreregex =
+    ```
+
+=== "/etc/fail2ban/jail.d/ntfy-ban.local"
+    ```
+    [ntfy-ban]
+    enabled = true
+    filter = ntfy-ban
+    action = iptables-multiport[name=ntfy-ban, port="http,https", protocol=tcp]
+    logpath = /var/log/ntfy/ban.log
+    maxretry = 1
+    findtime = 1m
+    bantime = 1h
+    ```
+
+
 ## IPv6 support
 ntfy fully supports IPv6, though there are a few things to keep in mind.
 
@@ -2328,6 +2396,10 @@ variable before running the `ntfy` command (e.g. `export NTFY_LISTEN_HTTP=:80`).
 | `visitor-topic-creation-limit-replenish`   | `NTFY_VISITOR_TOPIC_CREATION_LIMIT_REPLENISH`   | *duration*                                          | 1m                | Rate limiting: Rate at which the per-visitor topic-creation bucket is refilled (one new topic per x).                                                                                                                                   |
 | `visitor-prefix-bits-ipv4`                 | `NTFY_VISITOR_PREFIX_BITS_IPV4`                 | *number*                                            | 32                | Rate limiting: Number of bits to use for IPv4 visitor prefix, e.g. 24 for /24                                                                                                                                                           |
 | `visitor-prefix-bits-ipv6`                 | `NTFY_VISITOR_PREFIX_BITS_IPV6`                 | *number*                                            | 64                | Rate limiting: Number of bits to use for IPv6 visitor prefix, e.g. 48 for /48                                                                                                                                                           |
+| `ban-file`                                 | `NTFY_BAN_FILE`                                 | *filename*                                          | -                 | Abuse ban-feed: file confirmed abusive visitor IPs are appended to, for fail2ban to tail. Empty disables the feature. See [Banning bad actors](#banning-bad-actors-fail2ban)                                                             |
+| `ban-window`                               | `NTFY_BAN_WINDOW`                               | *duration*                                          | 10m               | Abuse ban-feed: rolling window over which weighted strikes are counted, per IP prefix                                                                                                                                                     |
+| `ban-threshold`                            | `NTFY_BAN_THRESHOLD`                            | *number*                                            | 100               | Abuse ban-feed: weighted strikes per `ban-window` before a prefix is written to `ban-file`                                                                                                                                             |
+| `ban-weights`                              | `NTFY_BAN_WEIGHTS`                              | *list of KEY:WEIGHT*                                | `42909:10 ... *:1`| Abuse ban-feed: per-code strike weights (exact code, family `429*`, or `*`; longest match wins; `0` exempts). See [Banning bad actors](#banning-bad-actors-fail2ban)                                                                     |
 | `web-root`                                 | `NTFY_WEB_ROOT`                                 | *path*, e.g. `/` or `/app`, or `disable`            | `/`               | Sets root of the web app (e.g. /, or /app), or disables it entirely (disable)                                                                                                                                                           |
 | `enable-signup`                            | `NTFY_ENABLE_SIGNUP`                            | *boolean* (`true` or `false`)                       | `false`           | Allows users to sign up via the web app, or API                                                                                                                                                                                         |
 | `enable-login`                             | `NTFY_ENABLE_LOGIN`                             | *boolean* (`true` or `false`)                       | `false`           | Allows users to log in via the web app, or API                                                                                                                                                                                          |

docs/contact.md

@@ -28,7 +28,7 @@ via the following channels:
 | Channel               | Contact                                             | Description                              |
 |-----------------------|-----------------------------------------------------|------------------------------------------|
 | **General Support**   | [support@mail.ntfy.sh](mailto:support@mail.ntfy.sh) | Direct email support for Pro subscribers |
-| **Billing Inquiries** | [billing@mail.ntfy.sh](mailto:support@mail.ntfy.sh) | Inquire about billing issues             |
+| **Billing Inquiries** | [billing@mail.ntfy.sh](mailto:billing@mail.ntfy.sh) | Inquire about billing issues             |
 | **Discord/Matrix**    | Mention your Pro status                             | Priority responses in community channels |
 
 Please include your ntfy.sh username when contacting support so we can verify your subscription status.
@@ -37,6 +37,11 @@ Please include your ntfy.sh username when contacting support so we can verify yo
 
 If you discover a security vulnerability, please report it responsibly via [security@mail.ntfy.sh](mailto:security@mail.ntfy.sh). See also: [SECURITY.md](https://github.com/binwiederhier/ntfy/blob/main/SECURITY.md).
 
+## Abuse reports
+
+To report spam, phishing, or other abuse of ntfy.sh, please email [abuse@mail.ntfy.sh](mailto:abuse@mail.ntfy.sh).
+Please include the topic name and any relevant message details so we can investigate.
+
 ## Other inquiries
 
 For questions about our [privacy policy](privacy.md), data handling, or to exercise your data rights 

docs/releases.md

@@ -2011,7 +2011,7 @@ and the [ntfy Android app](https://github.com/binwiederhier/ntfy-android/release
 
 **Features:**
 
-* Add an abuse ban-feed: when enabled via `ban-file`, ntfy tracks a weighted strike budget per visitor and appends abusive IPs to a file that fail2ban can tail and ban on sight (see `ban-file`, `ban-window`, `ban-threshold`, `ban-weights`)
+* Add an abuse ban-feed: when enabled via `ban-file`, ntfy tracks a weighted strike budget per visitor and appends abusive IPs to a file that fail2ban can tail and ban on sight (`ban-file`, `ban-window`, `ban-threshold`, `ban-weights`; see [ban-feed docs](config.md#ban-feed))
 
 ### ntfy Android v1.25.1 (UNRELEASED)
 

server/ban_test.go

@@ -0,0 +1,55 @@
+package server
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+	"heckel.io/ntfy/v2/ban"
+)
+
+// TestServer_BanFeed_WritesOffenderToFile is the end-to-end wiring test: a rejected request flows
+// through s.handle -> the error responder -> s.ban.Record, and once the offender's prefix
+// breaches, its ban line lands in the ban file (flushed on Close).
+func TestServer_BanFeed_WritesOffenderToFile(t *testing.T) {
+	banFile := filepath.Join(t.TempDir(), "ban.log")
+	conf := newTestConfig(t, "")
+	conf.BanFile = banFile
+	conf.BanWindow = time.Minute
+	conf.BanThreshold = 1                 // capacity 1: the 2nd rejection breaches
+	conf.BanWeights = ban.Weights{"*": 1} // any 4xx counts one strike
+	s := newTestServer(t, conf)
+	require.NotNil(t, s.ban)
+
+	// A delayed message with caching disabled is a deterministic 400 (errHTTPBadRequestDelayNoCache).
+	// request() sends from RemoteAddr 9.9.9.9.
+	reject := map[string]string{"Cache": "no", "In": "30 min"}
+	for i := 0; i < 3; i++ {
+		response := request(t, s, "PUT", "/mytopic", "", reject)
+		require.Equal(t, 400, response.Code)
+	}
+
+	// Writes are async; Close flushes the buffer. The offender's prefix must be in the feed exactly
+	// once (throttled to one line per window).
+	s.ban.Close()
+	data, err := os.ReadFile(banFile)
+	require.NoError(t, err)
+	lines := strings.Split(strings.TrimRight(string(data), "\n"), "\n")
+	require.Len(t, lines, 1)
+	require.Contains(t, lines[0], " 9.9.9.9 9.9.9.9/32 400 ") // <ip> <prefix> <http-code> <ntfy-code>
+}
+
+// TestServer_BanFeed_DisabledByDefault verifies the feature is off with no ban file: s.ban is
+// nil and the error path skips it (guarded), so a rejected request must not panic.
+func TestServer_BanFeed_DisabledByDefault(t *testing.T) {
+	conf := newTestConfig(t, "") // no BanFile
+	s := newTestServer(t, conf)
+	require.Nil(t, s.ban)
+
+	reject := map[string]string{"Cache": "no", "In": "30 min"}
+	response := request(t, s, "PUT", "/mytopic", "", reject) // guarded callsite, no Record
+	require.Equal(t, 400, response.Code)
+}

server/config.go

@@ -7,11 +7,10 @@ import (
 	"io/fs"
 	"net/netip"
 	"reflect"
-	"strconv"
-	"strings"
 	"text/template"
 	"time"
 
+	"heckel.io/ntfy/v2/ban"
 	"heckel.io/ntfy/v2/user"
 )
 
@@ -46,15 +45,25 @@ const (
 
 // Defines default abuse ban-feed settings (see BanFile, BanWindow, BanThreshold, BanWeights)
 const (
-	DefaultBanWindow    = time.Minute
+	DefaultBanWindow    = 10 * time.Minute
-	DefaultBanThreshold = 100 // Weighted strikes per BanWindow before a visitor is banned
+	DefaultBanThreshold = 100 // Weighted strikes per BanWindow before a prefix is banned
 )
 
-// DefaultBanWeights defines the per-code strike weights used when the ban-feed is enabled but no
+// DefaultBanWeights is the ban-feed's default per-code strike weights: the auth-failure flood bans fast,
-// explicit ban-weights are configured. Format is "KEY:WEIGHT" (see ParseBanWeights). The auth-failure
+// legit quota codes are exempt (weight 0), and everything else defaults to weight 1 (no "*"; see WeightFor).
-// flood is weighted heavily so it bans fast, the legit quota 429s are exempt (weight 0), and every
+var DefaultBanWeights = []string{
-// other rejection costs one strike via the "*" fallback.
+	banWeight(errHTTPTooManyRequestsLimitAuthFailure, 10),        // brute-force auth flood -> ban fast
-var DefaultBanWeights = []string{"42909:10", "42908:0", "42903:0", "42905:0", "42910:0", "*:1"}
+	banWeight(errHTTPTooManyRequestsLimitMessages, 0),            // daily message quota -> legit, exempt
+	banWeight(errHTTPTooManyRequestsLimitSubscriptions, 0),       // too many subscriptions -> legit, exempt
+	banWeight(errHTTPTooManyRequestsLimitAttachmentBandwidth, 0), // daily attachment bandwidth -> legit, exempt
+	banWeight(errHTTPTooManyRequestsLimitCalls, 0),               // daily phone-call quota -> legit, exempt
+}
+
+// banWeight formats a "CODE:WEIGHT" ban-feed default from an ntfy error, so the codes stay in sync with
+// the errHTTP definitions instead of being duplicated as string literals.
+func banWeight(err *errHTTP, weight int) string {
+	return fmt.Sprintf("%d:%d", err.Code, weight)
+}
 
 // Defines all global and per-visitor limits
 // - message size limit: the max number of bytes for a message
@@ -213,8 +222,8 @@ type Config struct {
 	WebPushExpiryWarningDuration         time.Duration
 	BanFile                              string        // Abuse ban-feed: file that fail2ban tails; empty string disables the feature
 	BanWindow                            time.Duration // Abuse ban-feed: rolling window over which weighted strikes are counted
-	BanThreshold                         int            // Abuse ban-feed: weighted strikes per window before a visitor is banned
+	BanThreshold                         int           // Abuse ban-feed: weighted strikes per window before a prefix is banned
-	BanWeights                           map[string]int // Abuse ban-feed: normalized code matcher -> strike weight (see ParseBanWeights, weightFor)
+	BanWeights                           ban.Weights   // Abuse ban-feed: code matcher -> strike weight (see ban.ParseWeights, ban.Weights.WeightFor)
 	BuildVersion                         string        // Injected by App
 	BuildDate                            string        // Injected by App
 	BuildCommit                          string        // Injected by App
@@ -345,83 +354,3 @@ func (c *Config) Hash() string {
 	}
 	return fmt.Sprintf("%x", sha256.Sum256([]byte(result)))
 }
-
-// ParseBanWeights turns a list like ["42909:10","403:2","42908:0","*:1"] into a map of normalized
-// matcher key -> strike weight for the abuse ban-feed's single weighted bucket. A key may be an exact
-// ntfy code ("42909"), a prefix family ("429*"), or "*". A bare 3-digit HTTP status ("403") is a
-// shorthand normalized to its family ("403*"), so operators can weight a whole status at once. Weights
-// must be integers >= 0; a weight of 0 is valid and means the code is exempt (never contributes to a
-// ban), which lets a "*" catch-all coexist with carved-out legit-quota codes. A malformed entry is
-// rejected so misconfiguration surfaces at startup rather than silently disabling bans.
-func ParseBanWeights(entries []string) (map[string]int, error) {
-	out := make(map[string]int, len(entries))
-	for _, entry := range entries {
-		key, weightStr, ok := strings.Cut(entry, ":")
-		if !ok {
-			return nil, fmt.Errorf("invalid ban-weight %q, want KEY:WEIGHT", entry)
-		}
-		weight, err := strconv.Atoi(strings.TrimSpace(weightStr))
-		if err != nil || weight < 0 {
-			return nil, fmt.Errorf("invalid ban-weight value in %q, want a non-negative integer", entry)
-		}
-		key = strings.TrimSpace(key)
-		if !validBanWeightKey(key) {
-			return nil, fmt.Errorf("invalid ban-weight key in %q, want %q, an ntfy code, an HTTP status, or a PREFIX*", entry, "*")
-		}
-		// A bare 3-digit HTTP status is shorthand for the whole family (e.g. "403" -> "403*").
-		if len(key) == 3 && isAllDigits(key) {
-			key += "*"
-		}
-		out[key] = weight
-	}
-	return out, nil
-}
-
-// validBanWeightKey reports whether key is a legal ban-weight matcher: "*", an all-digits code,
-// or an all-digits prefix followed by "*".
-func validBanWeightKey(key string) bool {
-	if key == "*" {
-		return true
-	}
-	digits := strings.TrimSuffix(key, "*")
-	return digits != "" && isAllDigits(digits)
-}
-
-func isAllDigits(s string) bool {
-	for _, r := range s {
-		if r < '0' || r > '9' {
-			return false
-		}
-	}
-	return s != ""
-}
-
-// weightFor returns the strike weight for a rejection's 5-digit ntfy code using the ban-weights
-// matcher, longest-match-wins. An exact key ("42909") matches with length len(code); a family key
-// ("429*") matches a code with that prefix, with length equal to the prefix; "*" matches everything
-// with length 0. The weight of the longest match is returned, or 0 if no key matches (which the
-// caller treats as "no strike").
-func (c *Config) weightFor(ntfyCode int) int {
-	code := strconv.Itoa(ntfyCode)
-	weight, bestLen, matched := 0, -1, false
-	for key, w := range c.BanWeights {
-		matchLen := -1
-		switch {
-		case key == "*":
-			matchLen = 0
-		case strings.HasSuffix(key, "*"):
-			if prefix := strings.TrimSuffix(key, "*"); strings.HasPrefix(code, prefix) {
-				matchLen = len(prefix)
-			}
-		case key == code:
-			matchLen = len(code)
-		}
-		if matchLen > bestLen {
-			weight, bestLen, matched = w, matchLen, true
-		}
-	}
-	if !matched {
-		return 0
-	}
-	return weight
-}

server/server.go

@@ -31,6 +31,7 @@ import (
 	"golang.org/x/sync/errgroup"
 	"heckel.io/ntfy/v2/action"
 	"heckel.io/ntfy/v2/attachment"
+	"heckel.io/ntfy/v2/ban"
 	"heckel.io/ntfy/v2/db"
 	"heckel.io/ntfy/v2/db/pg"
 	"heckel.io/ntfy/v2/log"
@@ -57,6 +58,7 @@ type Server struct {
 	mailer            mail.Sender
 	topics            map[string]*topic
 	visitors          map[string]*visitor // ip:<ip> or user:<user>
+	ban               *ban.Service        // Abuse ban-feed; nil when the feature is disabled (no ban file)
 	firebaseClient    *firebaseClient
 	twilio            *twilioClient
 	messages          int64                               // Total number of messages (persisted if messageCache enabled)
@@ -286,6 +288,17 @@ func New(conf *Config) (*Server, error) {
 		}
 		firebaseClient = newFirebaseClient(sender, auther)
 	}
+	var banner *ban.Service
+	if conf.BanFile != "" {
+		banner = ban.NewService(&ban.Config{
+			File:           conf.BanFile,
+			Window:         conf.BanWindow,
+			Threshold:      conf.BanThreshold,
+			Weights:        conf.BanWeights,
+			PrefixBitsIPv4: conf.VisitorPrefixBitsIPv4,
+			PrefixBitsIPv6: conf.VisitorPrefixBitsIPv6,
+		})
+	}
 	s := &Server{
 		config:          conf,
 		db:              pool,
@@ -295,6 +308,7 @@ func New(conf *Config) (*Server, error) {
 		firebaseClient:  firebaseClient,
 		twilio:          newTwilioClient(conf, userManager),
 		mailer:          sender,
+		ban:             banner,
 		topics:          topics,
 		userManager:     userManager,
 		messages:        messages,
@@ -450,6 +464,9 @@ func (s *Server) Stop() {
 		s.attachment.Close()
 	}
 	s.closeDatabases()
+	if s.ban != nil {
+		s.ban.Close()
+	}
 	if s.closeChan != nil {
 		close(s.closeChan)
 	}
@@ -538,13 +555,9 @@ func (s *Server) handleError(w http.ResponseWriter, r *http.Request, v *visitor,
 	w.Header().Set("Access-Control-Allow-Origin", s.config.AccessControlAllowOrigin) // CORS, allow cross-origin requests
 	w.WriteHeader(httpErr.HTTPCode)
 	io.WriteString(w, httpErr.JSON()+"\n")
-	if s.config.BanFile != "" {
+	if s.ban != nil {
-		// Abuse ban-feed: record against the OFFENDING request's IP, not v.ip. An account-keyed
-		// (tier'd) visitor is one object shared across all its source IPs, so v.ip is stale. The IP was
-		// already extracted in maybeAuthenticate and stashed in the request context (contextVisitorIP),
-		// so reuse it here rather than re-parsing headers.
 		if ip, err := fromContext[netip.Addr](r, contextVisitorIP); err == nil {
-			v.recordStatus(ip, httpErr.HTTPCode, httpErr.Code)
+			s.ban.Record(ip, httpErr.HTTPCode, httpErr.Code)
 		}
 	}
 }

server/server_middleware.go

@@ -13,7 +13,7 @@ const (
 	contextRateVisitor contextKey = iota + 2586
 	contextTopic
 	contextMatrixPushKey
-	contextVisitorIP // Client IP extracted in maybeAuthenticate; reused by the abuse ban-feed (see recordStatus)
+	contextVisitorIP // Client IP extracted in maybeAuthenticate; reused by the abuse ban-feed (see ban.Service.Record)
 )
 
 func (s *Server) limitRequests(next handleFunc) handleFunc {

server/server_test.go

@@ -5167,6 +5167,7 @@ func TestServer_BanFeed_RateLimitedIPBanned(t *testing.T) {
 		}
 	}
 	require.Greater(t, got429, 2)
+	s.ban.Close() // Writes are async (runWriteLoop); Close flushes the buffer before we read
 	data, err := os.ReadFile(banFile)
 	require.NoError(t, err)
 	require.Contains(t, string(data), "9.9.9.9 9.9.9.9/32 429 42901") // <ip> <prefix> <http> <ntfy-code>
@@ -5186,5 +5187,6 @@ func TestServer_BanFeed_SuccessfulRequestsNotBanned(t *testing.T) {
 		rr := request(t, s, "PUT", "/mytopic", fmt.Sprintf("m%d", i), nil)
 		require.Equal(t, 200, rr.Code)
 	}
+	s.ban.Close() // Flush any buffered bans (there should be none) before asserting no file
 	require.NoFileExists(t, banFile)
 }

server/visitor.go

@@ -4,7 +4,6 @@ import (
 	"fmt"
 	"math"
 	"net/netip"
-	"os"
 	"sync"
 	"time"
 
@@ -71,8 +70,6 @@ type visitor struct {
 	authLimiter          *rate.Limiter      // Limiter for incorrect login attempts, may be nil
 	firebase             time.Time          // Next allowed Firebase message
 	seen                 time.Time          // Last seen time of this visitor (needed for removal of stale visitors)
-	banLimiter           *rate.Limiter      // Abuse ban-feed: single weighted breach detector (rate.Limiter is concurrency-safe), nil when feature disabled
-	banEmit              *rate.Limiter      // Abuse ban-feed: throttles writes so a persistent offender only re-appears occasionally, nil when disabled
 	mu                   sync.RWMutex
 }
 
@@ -145,13 +142,6 @@ func newVisitor(conf *Config, messageCache *message.Cache, userManager *user.Man
 		accountLimiter:       nil, // Set in resetLimiters, may be nil
 		authLimiter:          nil, // Set in resetLimiters, may be nil
 	}
-	// Abuse ban-feed: only wire up per-visitor tracking when the feature is enabled. One weighted
-	// token bucket (capacity BanThreshold, refilled at BanThreshold/BanWindow per second) is the
-	// breach detector; banEmit throttles writes so the feed stays tiny.
-	if conf.BanFile != "" {
-		v.banLimiter = rate.NewLimiter(rate.Limit(float64(conf.BanThreshold)/conf.BanWindow.Seconds()), conf.BanThreshold)
-		v.banEmit = rate.NewLimiter(rate.Every(conf.BanWindow), 1)
-	}
 	v.resetLimitersNoLock(messages, emails, calls, false)
 	return v
 }
@@ -561,73 +551,6 @@ func dailyLimitToRate(limit int64) rate.Limit {
 	return rate.Limit(limit) * rate.Every(oneDay)
 }
 
-// recordStatus is called once per request with the offending request's IP and the final HTTP status
-// and ntfy code. Each rejection (4xx/5xx only) consumes weightFor(ntfyCode) tokens from the visitor's
-// single weighted ban bucket; a code weighted 0 (or one that matches no rule) is exempt and never
-// counts. When the bucket cannot cover a rejection the visitor has breached, and ip is appended to
-// the ban file that fail2ban tails (throttled via banEmit so a persistent offender only re-appears
-// occasionally). ip is passed in rather than read from v.ip because an account-keyed (tier'd) visitor
-// is shared across all its source IPs, so v.ip would be stale; we ban the address that breached.
-// This is on the hot path, so it no-ops immediately when the feature is disabled.
-func (v *visitor) recordStatus(ip netip.Addr, httpCode, ntfyCode int) {
-	if v.config.BanFile == "" {
-		return // Feature disabled
-	}
-	if httpCode < 400 {
-		return // Only rejections (4xx/5xx) count toward a ban; success and redirects never do
-	}
-	// Note: tier'd (account-keyed) visitors are NOT exempt. A persistent 429 stream is abusive
-	// regardless of the account behind it, and banning the offending IP is the right response --
-	// exactly what the pre-existing nginx-side jails already did. The legit case (a paid user merely
-	// hitting a plan/quota limit) is spared by the per-code weights instead: quota codes like 42908
-	// are weighted 0 below. Only request-limiter floods (42901, weight 1) actually accrue strikes.
-	weight := v.config.weightFor(ntfyCode)
-	if weight <= 0 {
-		return // Exempt (weight 0) or unmatched code: no strike
-	}
-	if v.banLimiter.AllowN(time.Now(), weight) {
-		return // The bucket covered this rejection; still within the strike budget
-	}
-	// Breached. Throttle so the feed stays tiny even if the offender keeps hammering.
-	if v.banEmit.Allow() {
-		writeBanLine(v.config.BanFile, ip, visitorPrefix(ip, v.config), httpCode, ntfyCode)
-	}
-}
-
-// banFileMu serializes appends to the ban file across all visitors (multiple visitors may breach
-// concurrently). The feed is pre-filtered, so write volume is low and a single mutex is plenty.
-var banFileMu sync.Mutex
-
-// writeBanLine appends a single line to the ban file:
-//
-//	<RFC3339-UTC-timestamp> <ip> <prefix> <http-code> <ntfy-code>
-//
-// e.g. "2026-07-18T20:56:32Z 1.2.3.4 1.2.3.4/32 429 42901". <prefix> is <ip> masked to the configured
-// rate-limiting prefix (VisitorPrefixBitsIPv4/IPv6); that is the unit fail2ban should ban, so an IPv6
-// client (which owns a whole /64) is banned as one, matching how ntfy rate-limits it. This exact
-// format is a contract that a fail2ban jail parses. Best-effort; any error is swallowed, because a
-// failure to write the ban feed must never fail the underlying request.
-func writeBanLine(path string, ip netip.Addr, prefix netip.Prefix, httpCode, ntfyCode int) {
-	banFileMu.Lock()
-	defer banFileMu.Unlock()
-	f, err := os.OpenFile(path, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
-	if err != nil {
-		return
-	}
-	defer f.Close()
-	fmt.Fprintf(f, "%s %s %s %d %d\n", time.Now().UTC().Format(time.RFC3339), ip.String(), prefix.String(), httpCode, ntfyCode)
-}
-
-// visitorPrefix masks ip to the configured rate-limiting prefix (VisitorPrefixBitsIPv4/IPv6), so the
-// ban feed reports the same unit ntfy rate-limits by -- e.g. a whole /64 for an IPv6 client -- rather
-// than a single address. fail2ban bans that prefix.
-func visitorPrefix(ip netip.Addr, conf *Config) netip.Prefix {
-	if ip.Is4() {
-		return netip.PrefixFrom(ip, conf.VisitorPrefixBitsIPv4).Masked()
-	}
-	return netip.PrefixFrom(ip, conf.VisitorPrefixBitsIPv6).Masked()
-}
-
 // visitorID returns a unique identifier for a visitor based on user or IP, using configurable prefix bits for IPv4/IPv6
 func visitorID(ip netip.Addr, u *user.User, conf *Config) string {
 	if u != nil && u.Tier != nil {

server/visitor_test.go

@@ -1,232 +0,0 @@
-package server
-
-import (
-	"net/netip"
-	"os"
-	"path/filepath"
-	"strings"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/require"
-	"heckel.io/ntfy/v2/user"
-)
-
-// newBanTestVisitor creates a visitor wired for ban-feed testing, with the given ban file,
-// weighted-bucket threshold, and weights, plus a 1-minute window (so the emit throttle only
-// fires once per test).
-func newBanTestVisitor(t *testing.T, banFile string, threshold int, weights map[string]int) *visitor {
-	conf := NewConfig()
-	conf.BanFile = banFile
-	conf.BanWindow = time.Minute
-	conf.BanThreshold = threshold
-	conf.BanWeights = weights
-	return newVisitor(conf, nil, nil, netip.MustParseAddr("1.2.3.4"), nil)
-}
-
-func readBanLines(t *testing.T, path string) []string {
-	data, err := os.ReadFile(path)
-	require.NoError(t, err)
-	return strings.Split(strings.TrimRight(string(data), "\n"), "\n")
-}
-
-func TestVisitor_RecordStatus_Weight2BansAtHalfThreshold(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	// Threshold 10, code weight 2 -> the budget covers exactly 5 hits, so the 6th breaches.
-	v := newBanTestVisitor(t, banFile, 10, map[string]int{"*": 2})
-	for i := 0; i < 5; i++ {
-		v.recordStatus(v.ip, 400, 40001)
-	}
-	require.NoFileExists(t, banFile) // 5 hits * weight 2 = 10 == budget, exactly at the limit, not over
-	v.recordStatus(v.ip, 400, 40001) // 6th hit cannot be covered -> breach
-	lines := readBanLines(t, banFile)
-	require.Len(t, lines, 1)
-	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 400 40001")) // <ip> <prefix> <http> <ntfy-code>
-}
-
-func TestVisitor_RecordStatus_Weight10BansFast(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	// Threshold 10, code weight 10 -> a single hit drains the whole budget, so the 2nd breaches.
-	v := newBanTestVisitor(t, banFile, 10, map[string]int{"42909": 10, "*": 1})
-	v.recordStatus(v.ip, 429, 42909)
-	require.NoFileExists(t, banFile)
-	v.recordStatus(v.ip, 429, 42909) // 2nd hit cannot be covered -> breach
-	lines := readBanLines(t, banFile)
-	require.Len(t, lines, 1)
-	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 429 42909"))
-}
-
-func TestVisitor_RecordStatus_Weight0NeverBans(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	// The legit-quota code is exempt (weight 0), so no number of hits ever bans.
-	v := newBanTestVisitor(t, banFile, 10, map[string]int{"42908": 0, "*": 1})
-	for i := 0; i < 100; i++ {
-		v.recordStatus(v.ip, 429, 42908)
-	}
-	require.NoFileExists(t, banFile)
-}
-
-func TestVisitor_RecordStatus_SingleBucketNoRelaxation(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	// One shared bucket: different codes draw down the SAME budget, so mixing them creates no extra
-	// headroom (unlike per-code buckets, which would relax the effective limit for a mixed offender).
-	v := newBanTestVisitor(t, banFile, 10, map[string]int{"403*": 2, "*": 1})
-	v.recordStatus(v.ip, 403, 40301) // weight 2 -> 8 left
-	v.recordStatus(v.ip, 403, 40301) // weight 2 -> 6 left
-	v.recordStatus(v.ip, 403, 40301) // weight 2 -> 4 left
-	require.NoFileExists(t, banFile)
-	for i := 0; i < 4; i++ {
-		v.recordStatus(v.ip, 400, 40001) // weight 1 each -> drains the remaining 4 -> 0 left
-	}
-	require.NoFileExists(t, banFile) // 6 + 4 = 10 == budget exactly, still not over
-	v.recordStatus(v.ip, 400, 40001) // one more cannot be covered -> breach
-	lines := readBanLines(t, banFile)
-	require.Len(t, lines, 1)
-}
-
-func TestVisitor_RecordStatus_ExactLineFormat(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	v := newBanTestVisitor(t, banFile, 1, map[string]int{"*": 1})
-	before := time.Now().UTC().Truncate(time.Second)
-	for i := 0; i < 3; i++ {
-		v.recordStatus(v.ip, 429, 42901)
-	}
-	after := time.Now().UTC()
-	lines := readBanLines(t, banFile)
-	require.Len(t, lines, 1)
-	parts := strings.Split(lines[0], " ")
-	require.Len(t, parts, 5) // "<timestamp> <ip> <prefix> <http-code> <ntfy-code>"
-	ts, err := time.Parse(time.RFC3339, parts[0])
-	require.NoError(t, err)
-	require.Equal(t, time.UTC, ts.Location())
-	require.False(t, ts.Before(before))
-	require.False(t, ts.After(after.Add(time.Second)))
-	require.Equal(t, "1.2.3.4", parts[1])    // full IP
-	require.Equal(t, "1.2.3.4/32", parts[2]) // masked to the default IPv4 prefix (/32)
-	require.Equal(t, "429", parts[3])        // HTTP status
-	require.Equal(t, "42901", parts[4])      // ntfy code
-}
-
-func TestVisitor_RecordStatus_IPv6MaskedToPrefix(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	conf := NewConfig()
-	conf.BanFile = banFile
-	conf.BanWindow = time.Minute
-	conf.BanThreshold = 1
-	conf.BanWeights = map[string]int{"*": 1}
-	v := newVisitor(conf, nil, nil, netip.MustParseAddr("2001:db8::abcd"), nil)
-	for i := 0; i < 3; i++ {
-		v.recordStatus(v.ip, 429, 42901)
-	}
-	parts := strings.Split(readBanLines(t, banFile)[0], " ")
-	require.Len(t, parts, 5)
-	require.Equal(t, "2001:db8::abcd", parts[1]) // full IPv6 address
-	require.Equal(t, "2001:db8::/64", parts[2])  // masked to the default IPv6 prefix (/64)
-}
-
-func TestVisitor_RecordStatus_TierUserAlsoBanned(t *testing.T) {
-	// A tier'd user who keeps hammering the request limiter (42901) is banned by IP like anyone
-	// else -- a persistent 429 stream is abusive regardless of the account behind it. The legit
-	// case (hitting a paid plan/quota limit) is spared by the per-code weights, not by a tier skip:
-	// codes like 42908 are weight 0. So tier does not exempt a visitor from the ban feed.
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	v := newBanTestVisitor(t, banFile, 1, map[string]int{"*": 1})
-	v.user = &user.User{ID: "u_test", Name: "test", Tier: &user.Tier{ID: "ti_test"}}
-	for i := 0; i < 10; i++ {
-		v.recordStatus(v.ip, 429, 42901)
-	}
-	lines := readBanLines(t, banFile)
-	require.Len(t, lines, 1)
-	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 429 42901"))
-}
-
-func TestVisitor_RecordStatus_BansOffendingIPNotVisitorIP(t *testing.T) {
-	// For an account-keyed (tier'd) visitor, one visitor object serves many source IPs and its
-	// stored v.ip is whichever IP created it first -- stale. The ban feed must write the IP of the
-	// request that actually breached, passed in per-call, not the visitor's stored IP.
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	v := newBanTestVisitor(t, banFile, 1, map[string]int{"*": 1}) // v.ip == 1.2.3.4
-	offender := netip.MustParseAddr("5.6.7.8")
-	for i := 0; i < 3; i++ {
-		v.recordStatus(offender, 429, 42901)
-	}
-	parts := strings.Split(readBanLines(t, banFile)[0], " ")
-	require.Equal(t, "5.6.7.8", parts[1])    // the offending request IP, not v.ip (1.2.3.4)
-	require.Equal(t, "5.6.7.8/32", parts[2]) // its prefix, not the visitor's stored-IP prefix
-}
-
-func TestVisitor_RecordStatus_DisabledWhenNoBanFile(t *testing.T) {
-	v := newBanTestVisitor(t, "", 10, map[string]int{"*": 1})
-	for i := 0; i < 5; i++ {
-		v.recordStatus(v.ip, 403, 40301) // Feature disabled: must be a no-op, must not panic
-	}
-	require.Nil(t, v.banEmit) // No throttle limiter is created when the feature is off
-}
-
-func TestVisitor_RecordStatus_Ignores2xx3xx(t *testing.T) {
-	banFile := filepath.Join(t.TempDir(), "ban.log")
-	v := newBanTestVisitor(t, banFile, 3, map[string]int{"*": 1})
-	// Success and redirects must never count toward a ban, even over the threshold -- otherwise a
-	// legit high-volume publisher (lots of 200s) would get banned.
-	for i := 0; i < 20; i++ {
-		v.recordStatus(v.ip, 200, 20000)
-		v.recordStatus(v.ip, 302, 30000)
-	}
-	require.NoFileExists(t, banFile)
-	// A 4xx over the same budget still gets written.
-	for i := 0; i < 5; i++ {
-		v.recordStatus(v.ip, 400, 40001)
-	}
-	lines := readBanLines(t, banFile)
-	require.Len(t, lines, 1)
-	require.True(t, strings.HasSuffix(lines[0], " 1.2.3.4 1.2.3.4/32 400 40001"))
-}
-
-func TestParseBanWeights(t *testing.T) {
-	// Exact codes, a bare 3-digit HTTP status (normalized to a family), an exempt code, and "*".
-	weights, err := ParseBanWeights([]string{"42909:10", "403:2", "42908:0", "*:1"})
-	require.NoError(t, err)
-	require.Equal(t, map[string]int{"42909": 10, "403*": 2, "42908": 0, "*": 1}, weights)
-
-	// A bare 3-digit HTTP status normalizes to its family.
-	weights, err = ParseBanWeights([]string{"429:5"})
-	require.NoError(t, err)
-	require.Equal(t, map[string]int{"429*": 5}, weights)
-
-	// An explicit family key stays as-is.
-	weights, err = ParseBanWeights([]string{"429*:5"})
-	require.NoError(t, err)
-	require.Equal(t, map[string]int{"429*": 5}, weights)
-
-	// Weight 0 is valid and means exempt.
-	weights, err = ParseBanWeights([]string{"42908:0"})
-	require.NoError(t, err)
-	require.Equal(t, map[string]int{"42908": 0}, weights)
-
-	_, err = ParseBanWeights([]string{"401"}) // Missing weight
-	require.Error(t, err)
-	_, err = ParseBanWeights([]string{"401:-1"}) // Negative weight
-	require.Error(t, err)
-	_, err = ParseBanWeights([]string{"401:abc"}) // Non-integer weight
-	require.Error(t, err)
-	_, err = ParseBanWeights([]string{"abc:10"}) // Non-numeric key
-	require.Error(t, err)
-	_, err = ParseBanWeights([]string{"4*3:10"}) // Star not at the end
-	require.Error(t, err)
-}
-
-func TestConfig_WeightFor(t *testing.T) {
-	conf := NewConfig()
-	weights, err := ParseBanWeights([]string{"42908:0", "42903:0", "42905:0", "42910:0", "42909:10", "429*:1", "403*:2", "4*:1", "5*:1"})
-	require.NoError(t, err)
-	conf.BanWeights = weights
-	// Longest-match-wins: exact 5-digit beats "429*" beats "4*" beats "*".
-	require.Equal(t, 0, conf.weightFor(42908))
-	require.Equal(t, 10, conf.weightFor(42909))
-	require.Equal(t, 1, conf.weightFor(42901))
-	require.Equal(t, 2, conf.weightFor(40311))
-	require.Equal(t, 1, conf.weightFor(40011))
-	require.Equal(t, 1, conf.weightFor(50312))
-	// A code that matches nothing (no "*" here) returns 0.
-	require.Equal(t, 0, conf.weightFor(30012))
-}