fix: make bucket CORS replication converge

Define a deterministic CORS replication register with durable tombstones, strict source timestamps, equal-time conflict ordering, full-state status, and heal convergence.

Serialize local and peer CORS transitions with a distributed namespace lock, validate canonical transport payloads, preserve initial-sync deletes, and fail closed on metadata errors.

Add adversarial, concurrent, restart, status, heal, signed admin-dispatch, protocol, and middleware coverage together with the reviewed site-replication design record.

Refs #75

Signed-off-by: Feng Ruohang <rh@vonng.com>

cmd/api-router.go

@@ -18,6 +18,7 @@
 package cmd
 
 import (
+	"errors"
 	"net"
 	"net/http"
 	"strconv"
@@ -657,6 +658,8 @@ func applyBucketCors(w http.ResponseWriter, r *http.Request, cfg *bktcors.Config
 	if origin == "" {
 		return false // not a CORS request
 	}
+	h := w.Header()
+	h.Add("Vary", "Origin")
 
 	isPreflight := r.Method == http.MethodOptions &&
 		r.Header.Get("Access-Control-Request-Method") != ""
@@ -664,45 +667,52 @@ func applyBucketCors(w http.ResponseWriter, r *http.Request, cfg *bktcors.Config
 	if isPreflight {
 		method := r.Header.Get("Access-Control-Request-Method")
 		reqHeaders := splitAndTrim(r.Header.Get("Access-Control-Request-Headers"))
-		rule, allowedHeaders, ok := cfg.MatchPreflight(origin, method, reqHeaders)
+		// A preflight response depends on all three request headers that
+		// determine the outcome, including when the request is rejected.
+		h.Add("Vary", "Access-Control-Request-Method")
+		h.Add("Vary", "Access-Control-Request-Headers")
+		rule, allowedOrigin, allowedHeaders, ok := cfg.MatchPreflight(origin, method, reqHeaders)
 		if !ok {
 			writeResponse(w, http.StatusForbidden, nil, mimeNone)
 			return true
 		}
-		h := w.Header()
+		setBucketCorsOriginHeaders(h, allowedOrigin, origin)
-		h.Set("Access-Control-Allow-Origin", origin)
 		h.Set("Access-Control-Allow-Methods", method)
 		if len(allowedHeaders) > 0 {
 			h.Set("Access-Control-Allow-Headers", strings.Join(allowedHeaders, ", "))
 		}
+		if len(rule.ExposeHeaders) > 0 {
+			h.Set("Access-Control-Expose-Headers", strings.Join(rule.ExposeHeaders, ", "))
+		}
 		if rule.MaxAgeSeconds > 0 {
 			h.Set("Access-Control-Max-Age", strconv.Itoa(rule.MaxAgeSeconds))
 		}
-		h.Set("Access-Control-Allow-Credentials", "true")
-		// A preflight response depends on all three request headers that
-		// determine the outcome, so cache variation must key on each of them.
-		h.Add("Vary", "Origin")
-		h.Add("Vary", "Access-Control-Request-Method")
-		h.Add("Vary", "Access-Control-Request-Headers")
 		writeResponse(w, http.StatusOK, nil, mimeNone)
 		return true
 	}
 
 	// Actual request: attach headers if the origin+method match.
-	rule, ok := cfg.MatchRule(origin, r.Method)
+	rule, allowedOrigin, ok := cfg.MatchRule(origin, r.Method)
 	if !ok {
 		return false // no matching rule → no CORS headers, continue normally
 	}
-	h := w.Header()
+	setBucketCorsOriginHeaders(h, allowedOrigin, origin)
-	h.Set("Access-Control-Allow-Origin", origin)
-	h.Set("Access-Control-Allow-Credentials", "true")
 	if len(rule.ExposeHeaders) > 0 {
 		h.Set("Access-Control-Expose-Headers", strings.Join(rule.ExposeHeaders, ", "))
 	}
-	h.Add("Vary", "Origin")
 	return false
 }
 
+func setBucketCorsOriginHeaders(h http.Header, allowedOrigin, requestOrigin string) {
+	if allowedOrigin == "*" {
+		h.Set("Access-Control-Allow-Origin", "*")
+		h.Del("Access-Control-Allow-Credentials")
+		return
+	}
+	h.Set("Access-Control-Allow-Origin", requestOrigin)
+	h.Set("Access-Control-Allow-Credentials", "true")
+}
+
 // splitAndTrim splits a comma-separated header list into trimmed, non-empty values.
 func splitAndTrim(s string) []string {
 	if s == "" {
@@ -766,13 +776,19 @@ func corsHandler(handler http.Handler) http.Handler {
 	globalCors := cors.New(opts).Handler(handler)
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if bucket, _ := request2BucketObjectName(r); bucket != "" && globalBucketMetadataSys != nil {
-			if cfg, _, err := globalBucketMetadataSys.GetCorsConfig(bucket); err == nil && cfg != nil {
+			cfg, _, err := globalBucketMetadataSys.GetCorsConfig(bucket)
+			if err == nil && cfg != nil {
 				if applyBucketCors(w, r, cfg) {
 					return
 				}
 				handler.ServeHTTP(w, r)
 				return
 			}
+			if err != nil && !errors.Is(err, errConfigNotFound) && r.Header.Get("Origin") != "" {
+				internalLogOnceIf(r.Context(), err, "bucket-cors-metadata")
+				handler.ServeHTTP(w, r)
+				return
+			}
 		}
 		globalCors.ServeHTTP(w, r)
 	})

cmd/bucket-cors-handlers.go

@@ -77,7 +77,7 @@ func (api objectAPIHandlers) PutBucketCorsHandler(w http.ResponseWriter, r *http
 		return
 	}
 
-	corsBytes, err := io.ReadAll(io.LimitReader(r.Body, r.ContentLength))
+	corsBytes, err := io.ReadAll(r.Body)
 	if err != nil {
 		writeErrorResponse(ctx, w, toAPIError(ctx, err), r.URL)
 		return
@@ -97,7 +97,7 @@ func (api objectAPIHandlers) PutBucketCorsHandler(w http.ResponseWriter, r *http
 		return
 	}
 
-	updatedAt, err := globalBucketMetadataSys.Update(ctx, bucket, bucketCorsConfig, corsBytes)
+	updatedAt, err := updateLocalBucketCORSMetadata(ctx, objAPI, bucket, corsBytes)
 	if err != nil {
 		writeErrorResponse(ctx, w, toAPIError(ctx, err), r.URL)
 		return
@@ -181,7 +181,7 @@ func (api objectAPIHandlers) DeleteBucketCorsHandler(w http.ResponseWriter, r *h
 		return
 	}
 
-	updatedAt, err := globalBucketMetadataSys.Delete(ctx, bucket, bucketCorsConfig)
+	updatedAt, err := updateLocalBucketCORSMetadata(ctx, objAPI, bucket, nil)
 	if err != nil {
 		writeErrorResponse(ctx, w, toAPIError(ctx, err), r.URL)
 		return

cmd/bucket-cors-handlers_test.go

@@ -98,6 +98,38 @@ func testBucketCorsHandlers(obj ObjectLayer, instanceType, bucketName string, ap
 		t.Fatalf("PUT malformed cors: expected 400, got %d", rec.Code)
 	}
 
+	// Missing Content-MD5 is rejected before the body is parsed.
+	req, err = newTestRequest(http.MethodPut, getBucketCorsURL("", bucketName),
+		int64(len(testCORSDoc)), bytes.NewReader([]byte(testCORSDoc)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header.Del("Content-Md5")
+	if err = signRequestV4(req, creds.AccessKey, creds.SecretKey); err != nil {
+		t.Fatal(err)
+	}
+	rec = httptest.NewRecorder()
+	apiRouter.ServeHTTP(rec, req)
+	if rec.Code != http.StatusBadRequest || !bytes.Contains(rec.Body.Bytes(), []byte("<Code>MissingContentMD5</Code>")) {
+		t.Fatalf("PUT cors without Content-MD5: expected MissingContentMD5, got %d: %s", rec.Code, rec.Body.String())
+	}
+
+	// A signed but incorrect Content-MD5 is rejected while reading the body.
+	req, err = newTestRequest(http.MethodPut, getBucketCorsURL("", bucketName),
+		int64(len(testCORSDoc)), bytes.NewReader([]byte(testCORSDoc)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header.Set("Content-Md5", getMD5HashBase64([]byte("different body")))
+	if err = signRequestV4(req, creds.AccessKey, creds.SecretKey); err != nil {
+		t.Fatal(err)
+	}
+	rec = httptest.NewRecorder()
+	apiRouter.ServeHTTP(rec, req)
+	if rec.Code != http.StatusBadRequest || !bytes.Contains(rec.Body.Bytes(), []byte("<Code>BadDigest</Code>")) {
+		t.Fatalf("PUT cors with bad Content-MD5: expected BadDigest, got %d: %s", rec.Code, rec.Body.String())
+	}
+
 	// Re-PUT the config so the store→GetCorsConfig→enforce seam below has
 	// something to enforce (the earlier DELETE removed it).
 	req, err = newTestSignedRequestV4(http.MethodPut, getBucketCorsURL("", bucketName),

cmd/bucket-cors-middleware_test.go

@@ -20,8 +20,10 @@ package cmd
 import (
 	"net/http"
 	"net/http/httptest"
+	"strings"
 	"testing"
 
+	"github.com/minio/minio/internal/auth"
 	"github.com/minio/minio/internal/bucket/cors"
 )
 
@@ -49,6 +51,28 @@ func TestPerBucketCorsPreflight(t *testing.T) {
 	if rec.Code != http.StatusOK {
 		t.Fatalf("preflight status = %d", rec.Code)
 	}
+	if got := rec.Header().Get("Access-Control-Expose-Headers"); got != "ETag" {
+		t.Fatalf("expose-headers = %q", got)
+	}
+	requireCorsOriginVary(t, rec.Header())
+}
+
+func TestPerBucketCorsActualRequestNoMatchVariesByOrigin(t *testing.T) {
+	cfg := &cors.Config{CORSRules: []cors.Rule{{
+		AllowedOrigins: []string{"https://allowed.example.com"},
+		AllowedMethods: []string{"GET"},
+	}}}
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, "/mybucket/obj", nil)
+	req.Header.Set("Origin", "https://denied.example.com")
+
+	if handled := applyBucketCors(rec, req, cfg); handled {
+		t.Fatal("actual request must continue when CORS does not match")
+	}
+	if got := rec.Header().Get("Access-Control-Allow-Origin"); got != "" {
+		t.Fatalf("allow-origin = %q", got)
+	}
+	requireCorsOriginVary(t, rec.Header())
 }
 
 func TestPerBucketCorsPreflightNoMatch(t *testing.T) {
@@ -68,6 +92,34 @@ func TestPerBucketCorsPreflightNoMatch(t *testing.T) {
 	if rec.Code != http.StatusForbidden {
 		t.Fatalf("expected 403 for disallowed origin, got %d", rec.Code)
 	}
+	requireCorsVary(t, rec.Header())
+}
+
+func TestPerBucketCorsPreflightWildcardOrigin(t *testing.T) {
+	cfg := &cors.Config{CORSRules: []cors.Rule{{
+		AllowedOrigins: []string{"*"},
+		AllowedMethods: []string{"GET"},
+		AllowedHeaders: []string{"*"},
+		ExposeHeaders:  []string{"ETag"},
+	}}}
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodOptions, "/mybucket/obj", nil)
+	req.Header.Set("Origin", "https://app.example.com")
+	req.Header.Set("Access-Control-Request-Method", http.MethodGet)
+
+	if handled := applyBucketCors(rec, req, cfg); !handled {
+		t.Fatal("expected preflight to be handled")
+	}
+	if got := rec.Header().Get("Access-Control-Allow-Origin"); got != "*" {
+		t.Fatalf("allow-origin = %q", got)
+	}
+	if got := rec.Header().Get("Access-Control-Allow-Credentials"); got != "" {
+		t.Fatalf("allow-credentials = %q", got)
+	}
+	if got := rec.Header().Get("Access-Control-Expose-Headers"); got != "ETag" {
+		t.Fatalf("expose-headers = %q", got)
+	}
+	requireCorsVary(t, rec.Header())
 }
 
 func TestPerBucketCorsActualRequest(t *testing.T) {
@@ -84,10 +136,120 @@ func TestPerBucketCorsActualRequest(t *testing.T) {
 	if handled {
 		t.Fatal("actual (non-preflight) request must not be terminated by CORS")
 	}
-	if got := rec.Header().Get("Access-Control-Allow-Origin"); got != "http://any.com" {
+	if got := rec.Header().Get("Access-Control-Allow-Origin"); got != "*" {
 		t.Fatalf("allow-origin = %q", got)
 	}
+	if got := rec.Header().Get("Access-Control-Allow-Credentials"); got != "" {
+		t.Fatalf("allow-credentials = %q", got)
+	}
 	if got := rec.Header().Get("Access-Control-Expose-Headers"); got != "ETag" {
 		t.Fatalf("expose-headers = %q", got)
 	}
 }
+
+func TestPerBucketCorsOriginPatternResponse(t *testing.T) {
+	cfg := &cors.Config{CORSRules: []cors.Rule{{
+		AllowedOrigins: []string{"https://app.example.com", "https://*", "*"},
+		AllowedMethods: []string{"GET"},
+	}}}
+
+	tests := []struct {
+		origin          string
+		wantOrigin      string
+		wantCredentials string
+	}{
+		{"https://app.example.com", "https://app.example.com", "true"},
+		{"https://other.example.com", "https://other.example.com", "true"},
+		{"http://other.example.com", "*", ""},
+	}
+
+	for _, tt := range tests {
+		rec := httptest.NewRecorder()
+		req := httptest.NewRequest(http.MethodGet, "/mybucket/obj", nil)
+		req.Header.Set("Origin", tt.origin)
+		if handled := applyBucketCors(rec, req, cfg); handled {
+			t.Fatal("actual request must not be terminated by CORS")
+		}
+		if got := rec.Header().Get("Access-Control-Allow-Origin"); got != tt.wantOrigin {
+			t.Fatalf("origin %q: allow-origin = %q, want %q", tt.origin, got, tt.wantOrigin)
+		}
+		if got := rec.Header().Get("Access-Control-Allow-Credentials"); got != tt.wantCredentials {
+			t.Fatalf("origin %q: allow-credentials = %q, want %q", tt.origin, got, tt.wantCredentials)
+		}
+	}
+}
+
+func TestBucketCorsMetadataErrorFailsClosed(t *testing.T) {
+	oldObjectAPI := newObjectLayerFn()
+	setObjectLayer(nil)
+	defer setObjectLayer(oldObjectAPI)
+
+	wrapped := corsHandler(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	for _, method := range []string{http.MethodGet, http.MethodOptions} {
+		rec := httptest.NewRecorder()
+		req := httptest.NewRequest(method, getGetObjectURL("", "cors-metadata-error", "object"), nil)
+		req.Header.Set("Origin", "https://app.example.com")
+		if method == http.MethodOptions {
+			req.Header.Set("Access-Control-Request-Method", http.MethodGet)
+		}
+		wrapped.ServeHTTP(rec, req)
+
+		if rec.Code != http.StatusNoContent {
+			t.Fatalf("%s status = %d, want %d", method, rec.Code, http.StatusNoContent)
+		}
+		if got := rec.Header().Get("Access-Control-Allow-Origin"); got != "" {
+			t.Fatalf("%s metadata error fell back to global allow-origin %q", method, got)
+		}
+		if got := rec.Header().Get("Access-Control-Allow-Credentials"); got != "" {
+			t.Fatalf("%s metadata error fell back to global credentials %q", method, got)
+		}
+	}
+}
+
+func TestBucketCorsNoConfigUsesGlobalFallback(t *testing.T) {
+	ExecObjectLayerAPITest(ExecObjectLayerAPITestArgs{
+		t:          t,
+		objAPITest: testBucketCorsNoConfigUsesGlobalFallback,
+		endpoints:  []string{"GetBucketCors"},
+	})
+}
+
+func testBucketCorsNoConfigUsesGlobalFallback(_ ObjectLayer, _ string, bucket string, _ http.Handler, _ auth.Credentials, t *testing.T) {
+	wrapped := corsHandler(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodGet, getGetObjectURL("", bucket, "object"), nil)
+	req.Header.Set("Origin", "https://app.example.com")
+	wrapped.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+	if got := rec.Header().Get("Access-Control-Allow-Origin"); got != "https://app.example.com" {
+		t.Fatalf("allow-origin = %q", got)
+	}
+	if got := rec.Header().Get("Access-Control-Allow-Credentials"); got != "true" {
+		t.Fatalf("allow-credentials = %q", got)
+	}
+}
+
+func requireCorsVary(t *testing.T, header http.Header) {
+	t.Helper()
+	values := strings.Join(header.Values("Vary"), ",")
+	for _, want := range []string{"Origin", "Access-Control-Request-Method", "Access-Control-Request-Headers"} {
+		if !strings.Contains(values, want) {
+			t.Fatalf("Vary = %q, missing %q", values, want)
+		}
+	}
+}
+
+func requireCorsOriginVary(t *testing.T, header http.Header) {
+	t.Helper()
+	if values := strings.Join(header.Values("Vary"), ","); !strings.Contains(values, "Origin") {
+		t.Fatalf("Vary = %q, missing Origin", values)
+	}
+}

cmd/site-replication.go

@@ -43,6 +43,7 @@ import (
 	"github.com/minio/minio-go/v7/pkg/replication"
 	"github.com/minio/minio-go/v7/pkg/set"
 	"github.com/minio/minio/internal/auth"
+	"github.com/minio/minio/internal/bucket/cors"
 	"github.com/minio/minio/internal/bucket/lifecycle"
 	sreplication "github.com/minio/minio/internal/bucket/replication"
 	"github.com/minio/minio/internal/logger"
@@ -1577,12 +1578,33 @@ func (c *SiteReplicationSys) PeerBucketMetadataUpdateHandler(ctx context.Context
 		return wrapSRErr(errInvalidArgument)
 	}
 
+	var corsConfigData []byte
+	if item.Cors != nil {
+		var err error
+		corsConfigData, err = decodeCORSReplicationPayload(item.Cors)
+		if err != nil {
+			return wrapSRErr(err)
+		}
+		if err = validateCORSReplicationPayload(corsConfigData); err != nil {
+			return wrapSRErr(err)
+		}
+		var unlock func()
+		ctx, unlock, err = lockBucketCORSMetadata(ctx, objectAPI, item.Bucket)
+		if err != nil {
+			return wrapSRErr(err)
+		}
+		defer unlock()
+	}
+
 	meta, err := readBucketMetadata(ctx, objectAPI, item.Bucket)
 	if err != nil {
 		return wrapSRErr(err)
 	}
 
 	if meta.Created.After(item.UpdatedAt) {
+		if item.Cors != nil {
+			replLogOnceIf(ctx, fmt.Errorf("ignoring CORS event for bucket %s from %v before bucket creation at %v", item.Bucket, item.UpdatedAt, meta.Created), "cors-event-before-bucket-creation-"+item.Bucket)
+		}
 		return nil
 	}
 
@@ -1633,12 +1655,12 @@ func (c *SiteReplicationSys) PeerBucketMetadataUpdateHandler(ctx context.Context
 	}
 
 	if item.Cors != nil {
-		configData, err := base64.StdEncoding.DecodeString(*item.Cors)
+		localState := newCORSReplicationState(meta.CorsConfigXML, meta.CorsConfigUpdatedAt)
-		if err != nil {
+		incoming := newCORSReplicationState(corsConfigData, item.UpdatedAt)
-			return wrapSRErr(err)
+		if compareCORSReplicationStates(localState, incoming) < 0 {
+			meta.CorsConfigXML = bytes.Clone(corsConfigData)
+			meta.CorsConfigUpdatedAt = item.UpdatedAt
 		}
-		meta.CorsConfigXML = configData
-		meta.CorsConfigUpdatedAt = item.UpdatedAt
 	}
 
 	return globalBucketMetadataSys.save(ctx, meta)
@@ -1758,30 +1780,222 @@ func (c *SiteReplicationSys) PeerBucketSSEConfigHandler(ctx context.Context, buc
 	return nil
 }
 
-// PeerBucketCorsConfigHandler - copies/deletes CORS config to local cluster.
+type corsReplicationStateKind uint8
-func (c *SiteReplicationSys) PeerBucketCorsConfigHandler(ctx context.Context, bucket string, corsConfig *string, updatedAt time.Time) error {
-	// skip overwrite if local update is newer than peer update.
-	if !updatedAt.IsZero() {
-		if _, updateTm, err := globalBucketMetadataSys.GetCorsConfig(bucket); err == nil && updateTm.After(updatedAt) {
-			return nil
-		}
-	}
 
-	if corsConfig != nil {
+const (
-		configData, err := base64.StdEncoding.DecodeString(*corsConfig)
+	corsReplicationBaseline corsReplicationStateKind = iota
-		if err != nil {
+	corsReplicationLive
-			return wrapSRErr(err)
+	corsReplicationTombstone
-		}
+)
-		_, err = globalBucketMetadataSys.Update(ctx, bucket, bucketCorsConfig, configData)
+
-		if err != nil {
+type corsReplicationState struct {
-			return wrapSRErr(err)
+	kind      corsReplicationStateKind
-		}
+	payload   []byte
+	updatedAt time.Time
+}
+
+func newCORSReplicationState(payload []byte, updatedAt time.Time) corsReplicationState {
+	state := corsReplicationState{updatedAt: updatedAt.UTC()}
+	switch {
+	case len(payload) > 0:
+		state.kind = corsReplicationLive
+		state.payload = bytes.Clone(payload)
+	case updatedAt.IsZero():
+		state.kind = corsReplicationBaseline
+	default:
+		state.kind = corsReplicationTombstone
+	}
+	return state
+}
+
+func compareCORSReplicationStates(a, b corsReplicationState) int {
+	switch {
+	case a.updatedAt.Before(b.updatedAt):
+		return -1
+	case a.updatedAt.After(b.updatedAt):
+		return 1
+	case a.kind < b.kind:
+		return -1
+	case a.kind > b.kind:
+		return 1
+	case a.kind == corsReplicationLive:
+		return bytes.Compare(a.payload, b.payload)
+	default:
+		return 0
+	}
+}
+
+func equalCORSReplicationStates(a, b corsReplicationState) bool {
+	return compareCORSReplicationStates(a, b) == 0
+}
+
+func decodeCORSReplicationPayload(encoded *string) ([]byte, error) {
+	if encoded == nil {
+		return nil, nil
+	}
+	payload, err := base64.StdEncoding.Strict().DecodeString(*encoded)
+	if err != nil {
+		return nil, fmt.Errorf("invalid CORS replication payload: %w", err)
+	}
+	if len(payload) == 0 || base64.StdEncoding.EncodeToString(payload) != *encoded {
+		return nil, fmt.Errorf("invalid CORS replication payload: %w", errInvalidArgument)
+	}
+	return payload, nil
+}
+
+func validateCORSReplicationPayload(payload []byte) error {
+	if payload == nil {
 		return nil
 	}
-
+	config, err := cors.ParseBucketCorsConfig(bytes.NewReader(payload))
-	// Delete cors config
-	_, err := globalBucketMetadataSys.Delete(ctx, bucket, bucketCorsConfig)
 	if err != nil {
+		return fmt.Errorf("invalid CORS replication payload: %w", errInvalidArgument)
+	}
+	if err = config.Validate(); err != nil {
+		return fmt.Errorf("invalid CORS replication payload: %w: %v", errInvalidArgument, err)
+	}
+	return nil
+}
+
+func corsReplicationStateFromInfo(info madmin.SRBucketInfo) (corsReplicationState, error) {
+	payload, err := decodeCORSReplicationPayload(info.CorsConfig)
+	if err != nil {
+		return corsReplicationState{}, err
+	}
+	if err = validateCORSReplicationPayload(payload); err != nil {
+		return corsReplicationState{}, err
+	}
+	if info.CorsConfig != nil && info.CorsConfigUpdatedAt.IsZero() {
+		return corsReplicationState{}, fmt.Errorf("live CORS replication payload has no source timestamp: %w", errInvalidArgument)
+	}
+	return newCORSReplicationState(payload, info.CorsConfigUpdatedAt), nil
+}
+
+func areCORSReplicationStatesEqual(sites []srBucketMetaInfo) bool {
+	if len(sites) == 0 {
+		return true
+	}
+	reference, err := corsReplicationStateFromInfo(sites[0].SRBucketInfo)
+	if err != nil {
+		return false
+	}
+	for _, site := range sites[1:] {
+		state, err := corsReplicationStateFromInfo(site.SRBucketInfo)
+		if err != nil || !equalCORSReplicationStates(reference, state) {
+			return false
+		}
+	}
+	return true
+}
+
+func (s corsReplicationState) encodedPayload() *string {
+	if s.kind != corsReplicationLive {
+		return nil
+	}
+	encoded := base64.StdEncoding.EncodeToString(s.payload)
+	return &encoded
+}
+
+func newBucketCORSReplicationEvent(bucket string, meta BucketMetadata) (madmin.SRBucketMeta, bool) {
+	if meta.CorsConfigUpdatedAt.IsZero() {
+		return madmin.SRBucketMeta{}, false
+	}
+	return madmin.SRBucketMeta{
+		Type:      madmin.SRBucketMetaTypeCorsConfig,
+		Bucket:    bucket,
+		Cors:      newCORSReplicationState(meta.CorsConfigXML, meta.CorsConfigUpdatedAt).encodedPayload(),
+		UpdatedAt: meta.CorsConfigUpdatedAt,
+	}, true
+}
+
+func lockBucketCORSMetadata(ctx context.Context, objectAPI ObjectLayer, bucket string) (context.Context, func(), error) {
+	// The lock name is deliberately different from .metadata.bin. Saving the
+	// metadata locks that object internally, and namespace locks are not
+	// re-entrant.
+	lock := objectAPI.NewNSLock(minioMetaBucket, pathJoin(bucketMetaPrefix, bucket, "cors-config.lock"))
+	lkctx, err := lock.GetLock(ctx, globalOperationTimeout)
+	if err != nil {
+		return nil, nil, err
+	}
+	return lkctx.Context(), func() { lock.Unlock(lkctx) }, nil
+}
+
+func updateLocalBucketCORSMetadata(ctx context.Context, objectAPI ObjectLayer, bucket string, configData []byte) (time.Time, error) {
+	return applyBucketCORSMetadata(ctx, objectAPI, bucket, configData, time.Time{}, true)
+}
+
+func applyBucketCORSMetadata(ctx context.Context, objectAPI ObjectLayer, bucket string, configData []byte, sourceUpdatedAt time.Time, local bool) (time.Time, error) {
+	if bucket == "" || (configData != nil && len(configData) == 0) {
+		return time.Time{}, errInvalidArgument
+	}
+	if err := validateCORSReplicationPayload(configData); err != nil {
+		return time.Time{}, err
+	}
+
+	ctx, unlock, err := lockBucketCORSMetadata(ctx, objectAPI, bucket)
+	if err != nil {
+		return time.Time{}, err
+	}
+	defer unlock()
+
+	var meta BucketMetadata
+	if local {
+		meta, err = loadBucketMetadataParse(ctx, objectAPI, bucket, true)
+	} else {
+		meta, err = readBucketMetadata(ctx, objectAPI, bucket)
+	}
+	if err != nil {
+		return time.Time{}, err
+	}
+
+	localState := newCORSReplicationState(meta.CorsConfigXML, meta.CorsConfigUpdatedAt)
+	updatedAt := sourceUpdatedAt.UTC()
+	if local {
+		updatedAt = UTCNow()
+		floor := meta.Created
+		if localState.updatedAt.After(floor) {
+			floor = localState.updatedAt
+		}
+		if !updatedAt.After(floor) {
+			updatedAt = floor.Add(time.Nanosecond)
+		}
+	} else {
+		// CreatedAt is the bucket-lineage floor: an event from an older
+		// incarnation of the bucket must not change the current one.
+		if updatedAt.Before(meta.Created) {
+			replLogOnceIf(ctx, fmt.Errorf("ignoring CORS event for bucket %s from %v before bucket creation at %v", bucket, updatedAt, meta.Created), "cors-event-before-bucket-creation-"+bucket)
+			return localState.updatedAt, nil
+		}
+		incoming := newCORSReplicationState(configData, updatedAt)
+		if compareCORSReplicationStates(localState, incoming) >= 0 {
+			return localState.updatedAt, nil
+		}
+	}
+
+	meta.CorsConfigXML = bytes.Clone(configData)
+	meta.CorsConfigUpdatedAt = updatedAt
+	if err = globalBucketMetadataSys.save(ctx, meta); err != nil {
+		return time.Time{}, err
+	}
+	return updatedAt, nil
+}
+
+// PeerBucketCorsConfigHandler - copies/deletes CORS config to local cluster.
+func (c *SiteReplicationSys) PeerBucketCorsConfigHandler(ctx context.Context, bucket string, corsConfig *string, updatedAt time.Time) error {
+	objectAPI := newObjectLayerFn()
+	if objectAPI == nil {
+		return errSRObjectLayerNotReady
+	}
+
+	if bucket == "" || updatedAt.IsZero() {
+		return wrapSRErr(errInvalidArgument)
+	}
+
+	configData, err := decodeCORSReplicationPayload(corsConfig)
+	if err != nil {
+		return wrapSRErr(err)
+	}
+	if _, err = applyBucketCORSMetadata(ctx, objectAPI, bucket, configData, updatedAt, false); err != nil {
 		return wrapSRErr(err)
 	}
 	return nil
@@ -1989,15 +2203,8 @@ func (c *SiteReplicationSys) syncToAllPeers(ctx context.Context, addOpts madmin.
 		}
 
 		// Replicate existing bucket CORS settings
-		corsConfigData, tm := meta.CorsConfigXML, meta.CorsConfigUpdatedAt
+		if corsEvent, ok := newBucketCORSReplicationEvent(bucket, meta); ok {
-		if len(corsConfigData) > 0 {
+			err = c.BucketMetaHook(ctx, corsEvent)
-			corsConfigStr := base64.StdEncoding.EncodeToString(corsConfigData)
-			err = c.BucketMetaHook(ctx, madmin.SRBucketMeta{
-				Type:      madmin.SRBucketMetaTypeCorsConfig,
-				Bucket:    bucket,
-				Cors:      &corsConfigStr,
-				UpdatedAt: tm,
-			})
 			if err != nil {
 				return errSRBucketMetaError(err)
 			}
@@ -3198,7 +3405,6 @@ func (c *SiteReplicationSys) siteReplicationStatus(ctx context.Context, objAPI O
 			replCfgs := make([]*sreplication.Config, numSites)
 			quotaCfgs := make([]*madmin.BucketQuota, numSites)
 			sseCfgSet := set.NewStringSet()
-			corsCfgSet := set.NewStringSet()
 			versionCfgSet := set.NewStringSet()
 			var tagCount, olockCfgCount, sseCfgCount, corsCfgCount, versionCfgCount int
 			for i, s := range slc {
@@ -3272,13 +3478,8 @@ func (c *SiteReplicationSys) siteReplicationStatus(ctx context.Context, objAPI O
 					}
 				}
 				if s.CorsConfig != nil {
-					configData, err := base64.StdEncoding.DecodeString(*s.CorsConfig)
+					if _, err := decodeCORSReplicationPayload(s.CorsConfig); err == nil {
-					if err != nil {
+						corsCfgCount++
-						continue
-					}
-					corsCfgCount++
-					if !corsCfgSet.Contains(string(configData)) {
-						corsCfgSet.Add(string(configData))
 					}
 				}
 				ss, ok := info.StatsSummary[s.DeploymentID]
@@ -3299,7 +3500,7 @@ func (c *SiteReplicationSys) siteReplicationStatus(ctx context.Context, objAPI O
 				if sseCfgCount > 0 {
 					ss.TotalSSEConfigCount++
 				}
-				if corsCfgCount > 0 {
+				if s.CorsConfig != nil {
 					ss.TotalCorsConfigCount++
 				}
 				if versionCfgCount > 0 {
@@ -3313,7 +3514,7 @@ func (c *SiteReplicationSys) siteReplicationStatus(ctx context.Context, objAPI O
 			tagMismatch := !isReplicated(tagCount, numSites, tagSet)
 			olockCfgMismatch := !isReplicated(olockCfgCount, numSites, olockConfigSet)
 			sseCfgMismatch := !isReplicated(sseCfgCount, numSites, sseCfgSet)
-			corsCfgMismatch := !isReplicated(corsCfgCount, numSites, corsCfgSet)
+			corsCfgMismatch := !areCORSReplicationStatesEqual(slc)
 			versionCfgMismatch := !isReplicated(versionCfgCount, numSites, versionCfgSet)
 			policyMismatch := !isBktPolicyReplicated(numSites, policies)
 			replCfgMismatch := !isBktReplCfgReplicated(numSites, replCfgs)
@@ -3783,10 +3984,10 @@ func (c *SiteReplicationSys) SiteReplicationMetaInfo(ctx context.Context, objAPI
 				bms.SSEConfigUpdatedAt = meta.EncryptionConfigUpdatedAt
 			}
 
+			bms.CorsConfigUpdatedAt = meta.CorsConfigUpdatedAt
 			if len(meta.CorsConfigXML) > 0 {
 				corsConfigStr := base64.StdEncoding.EncodeToString(meta.CorsConfigXML)
 				bms.CorsConfig = &corsConfigStr
-				bms.CorsConfigUpdatedAt = meta.CorsConfigUpdatedAt
 			}
 
 			if len(meta.ReplicationConfigXML) > 0 {
@@ -4997,62 +5198,45 @@ func (c *SiteReplicationSys) healSSEMetadata(ctx context.Context, objAPI ObjectL
 	return nil
 }
 
+func latestCORSConfig(bs map[string]srBucketStatsSummary) (latestID string, latest corsReplicationState, ok bool) {
+	for dID, status := range bs {
+		state, err := corsReplicationStateFromInfo(status.meta.SRBucketInfo)
+		if err != nil || state.kind == corsReplicationBaseline {
+			continue
+		}
+		cmp := compareCORSReplicationStates(latest, state)
+		if !ok || cmp < 0 || (cmp == 0 && dID > latestID) {
+			latestID = dID
+			latest = state
+			ok = true
+		}
+	}
+	return latestID, latest, ok
+}
+
 func (c *SiteReplicationSys) healCORSMetadata(ctx context.Context, objAPI ObjectLayer, bucket string, info srStatusInfo) error {
 	c.RLock()
 	defer c.RUnlock()
 	if !c.enabled {
 		return nil
 	}
-	var (
-		latestID, latestPeerName string
-		lastUpdate               time.Time
-		latestCorsConfig         *string
-	)
 
 	bs := info.BucketStats[bucket]
-	for dID, ss := range bs {
+	latestID, latestState, ok := latestCORSConfig(bs)
-		if lastUpdate.IsZero() {
+	if !ok {
-			lastUpdate = ss.meta.CorsConfigUpdatedAt
+		return nil
-			latestID = dID
-			latestCorsConfig = ss.meta.CorsConfig
-		}
-		// avoid considering just created buckets as latest. Perhaps this site
-		// just joined cluster replication and yet to be sync'd
-		if ss.meta.CreatedAt.Equal(ss.meta.CorsConfigUpdatedAt) {
-			continue
-		}
-		if ss.meta.CorsConfigUpdatedAt.After(lastUpdate) {
-			lastUpdate = ss.meta.CorsConfigUpdatedAt
-			latestID = dID
-			latestCorsConfig = ss.meta.CorsConfig
-		}
 	}
 
-	latestPeerName = info.Sites[latestID].Name
+	latestPeerName := info.Sites[latestID].Name
-	var latestCorsConfigBytes []byte
+	latestCorsConfig := latestState.encodedPayload()
-	var err error
-	if latestCorsConfig != nil {
-		latestCorsConfigBytes, err = base64.StdEncoding.DecodeString(*latestCorsConfig)
-		if err != nil {
-			return err
-		}
-	}
 
 	for dID, bStatus := range bs {
-		if !bStatus.CorsCfgMismatch {
+		currentState, err := corsReplicationStateFromInfo(bStatus.meta.SRBucketInfo)
-			continue
+		if err == nil && equalCORSReplicationStates(latestState, currentState) {
-		}
-		if isBucketMetadataEqual(latestCorsConfig, bStatus.meta.CorsConfig) {
 			continue
 		}
 		if dID == globalDeploymentID() {
-			if latestCorsConfig == nil {
+			if err := c.PeerBucketCorsConfigHandler(ctx, bucket, latestCorsConfig, latestState.updatedAt); err != nil {
-				if _, err := globalBucketMetadataSys.Delete(ctx, bucket, bucketCorsConfig); err != nil {
-					replLogIf(ctx, fmt.Errorf("Unable to heal CORS metadata from peer site %s : %w", latestPeerName, err))
-				}
-				continue
-			}
-			if _, err := globalBucketMetadataSys.Update(ctx, bucket, bucketCorsConfig, latestCorsConfigBytes); err != nil {
 				replLogIf(ctx, fmt.Errorf("Unable to heal CORS metadata from peer site %s : %w", latestPeerName, err))
 			}
 			continue
@@ -5067,7 +5251,7 @@ func (c *SiteReplicationSys) healCORSMetadata(ctx context.Context, objAPI Object
 			Type:      madmin.SRBucketMetaTypeCorsConfig,
 			Bucket:    bucket,
 			Cors:      latestCorsConfig,
-			UpdatedAt: lastUpdate,
+			UpdatedAt: latestState.updatedAt,
 		})
 		if err != nil {
 			replLogIf(ctx, c.annotatePeerErr(peerName, replicateBucketMetadata,
@@ -5394,7 +5578,7 @@ func isBucketMetadataEqual(one, two *string) bool {
 	case one == nil || two == nil:
 		return false
 	default:
-		return strings.EqualFold(*one, *two)
+		return *one == *two
 	}
 }
 

docs/site-replication/CORS-LWW-DESIGN.md

@@ -0,0 +1,539 @@
+# Per-Bucket CORS Site-Replication Convergence Design
+
+## Status
+
+- Issue: [pgsty/silo#75](https://github.com/pgsty/silo/issues/75)
+- Baseline: `e4e3007da6d7d1198a6a050e34f84566d40a9654`
+- Working branch: `codex/issue-75-cors-hardening`
+- Decision: CORS-specific deterministic last-writer-wins register, described below
+- Implementation state: implemented and locally verified; uncommitted and unpushed
+- Release state: not released; remote CI, merge, tag, package, and image gates remain separate
+- Final design/implementation review: Claude Code Opus 5 Max found no P0/P1 and judged the implementation GO; its mandatory documentation corrections are incorporated here
+
+This document defines the replication state, ordering, persistence, status,
+healing, concurrency, compatibility, and test contract for per-bucket CORS.
+It is an implementation design record, not public upgrade or rollback guidance.
+Public operator documentation belongs in the separate `silo.pgsty.com`
+repository.
+
+## Scope
+
+This design covers the current-version CORS path:
+
+```text
+PutBucketCors / DeleteBucketCors
+    -> persist local CORS state
+    -> BucketMetaHook
+    -> madmin SRBucketMeta transport
+    -> SRPeerReplicateBucketItem dispatch
+    -> PeerBucketCorsConfigHandler
+    -> SiteReplicationMetaInfo
+    -> siteReplicationStatus
+    -> latestCORSConfig
+    -> healCORSMetadata
+```
+
+It also covers retry, duplicate delivery, reordering, equal timestamps,
+initial site sync, missed DELETE recovery, cache reload, process restart, and
+concurrent CORS mutations on different nodes of one cluster.
+
+The following are deliberately out of scope:
+
+- redesigning the replication semantics of policy, tags, SSE, quota,
+  versioning, or Object Lock;
+- eliminating lost updates between different bucket-metadata types that all
+  rewrite `.metadata.bin`; this inherited problem is tracked by
+  [pgsty/silo#77](https://github.com/pgsty/silo/issues/77);
+- mixed-version support that permits CORS writes before every site runs a
+  CORS-aware binary;
+- public downgrade, rollback, and global-fallback documentation;
+- Console UI for bucket CORS.
+
+### Adjacent issue-75 changes in the same candidate
+
+The current dirty issue-75 candidate also contains CORS work outside the LWW
+register itself:
+
+- stricter `cors.Config.Validate()` rules for empty origins and unsupported
+  wildcard forms;
+- matcher signature and response-selection changes needed to distinguish a
+  literal `*` origin from a patterned match;
+- fail-closed metadata-error handling in the HTTP middleware;
+- preflight expose headers and complete `Vary` behavior; and
+- HTTP protocol negative tests.
+
+Those changes share the same CORS release gate and are present in the reviewed
+diff, but they are not part of the replication conflict key or join algorithm.
+This document describes them only where they constrain replication validation
+or the final verification boundary.
+
+## Confirmed Failures in the Pre-Fix Candidate
+
+The pre-fix issue-75 candidate had four independently reproduced convergence
+defects:
+
+1. Heal compared only payloads. If two sites stored identical payload bytes
+   with different source timestamps, heal skipped the older site. The sites
+   retained different ordering barriers and could disagree on a later delayed
+   event.
+2. `isBucketMetadataEqual` used `strings.EqualFold` for base64. For example,
+   `QQ==` and `qQ==` decode to different bytes but compared equal.
+3. Status derived `CorsCfgMismatch` from live payload count and payload set.
+   It did not include source timestamp or tombstone state, so it could report
+   divergent sites as converged and suppress healing.
+4. Equal-timestamp conflicting events had no stable tie-breaker. Peer apply
+   accepted whichever event arrived last, while heal selected whichever map
+   entry happened to be visited first.
+
+Two additional correctness requirements followed from the state model:
+
+- the read, compare, and save transition must be atomic across nodes in one
+  cluster; and
+- a successful local PUT or DELETE must advance beyond an already stored
+  future source timestamp instead of moving the local barrier backwards.
+
+## Constraints
+
+The minimum fix must satisfy these constraints:
+
+- preserve `madmin.SRBucketMeta` and `madmin.SRBucketInfo` wire schemas;
+- preserve the exact source `UpdatedAt` on peer apply and heal;
+- distinguish a never-configured bucket from a persisted deletion;
+- converge without relying on event arrival order, map iteration order, or a
+  particular site being the healer;
+- serialize CORS-versus-CORS transitions cluster-wide without introducing a
+  broad bucket-metadata redesign;
+- reject malformed replication payloads before persistence;
+- remain idempotent under retry and initial-sync replay;
+- keep the replication state-machine change limited to CORS except for the
+  directly shared base64 equality bug; do not infer that the same dirty
+  candidate contains no adjacent CORS protocol or middleware changes.
+
+## State Model
+
+For one bucket lineage, the persisted CORS state is:
+
+```text
+State = (Payload, SourceUpdatedAt)
+```
+
+`BucketMetadata.Created` is not part of the conflict key. It is the lineage
+floor used to reject an event from an older incarnation of the bucket.
+
+### State kinds
+
+| Kind | Payload | `CorsConfigUpdatedAt` | Meaning |
+| --- | --- | --- | --- |
+| Baseline | nil | zero | CORS has never been configured for this bucket lineage |
+| Live | non-empty XML bytes | non-zero | A live per-bucket CORS configuration |
+| Tombstone | nil | non-zero | CORS was explicitly deleted at the source timestamp |
+
+The baseline uses a zero timestamp deliberately. Defaulting a missing CORS
+timestamp to `CreatedAt` would make classification depend on two values that
+can be obtained from different cache/disk snapshots. It would also make a
+never-configured state indistinguishable from a deletion at bucket creation.
+
+Per-bucket CORS and `CorsConfigUpdatedAt` were introduced together, so there
+is no released legacy live-CORS state that requires synthesizing a timestamp.
+
+### Wire canonicalization
+
+A non-nil wire payload must satisfy all of the following:
+
+1. strict standard base64 decoding succeeds;
+2. re-encoding the decoded bytes produces exactly the received string;
+3. the decoded payload is non-empty;
+4. CORS XML parsing succeeds; and
+5. `cors.Config.Validate()` succeeds.
+
+The canonical re-encode check rejects ignored newlines and alternate textual
+representations. Equality is therefore equality of decoded bytes, with exact
+base64 string equality remaining safe for the shared metadata helper.
+
+An invalid wire value is not a candidate winner and is never propagated.
+Peer apply rejects it before any metadata write.
+
+## Deterministic Ordering
+
+States use the following total order:
+
+```text
+1. SourceUpdatedAt
+2. Kind: baseline < live < tombstone
+3. For live/live ties: lexicographic decoded payload bytes
+```
+
+The greater state wins.
+
+Consequences:
+
+- a newer source event wins regardless of arrival order;
+- the same payload with a newer timestamp is a greater state and advances the
+  ordering barrier;
+- a DELETE wins an equal-timestamp PUT/DELETE conflict;
+- two equal-timestamp live payloads choose the same bytewise winner at every
+  site;
+- an exact duplicate is equal and therefore a no-op;
+- retry, reordering, and duplicate delivery cannot move local state backward.
+
+The live-payload tie-breaker is not intended to identify the human's temporal
+intent. It supplies the deterministic result required when the timestamp has
+already failed to distinguish two writes.
+
+## Why No Source-Site Tie-Breaker
+
+The rejected source-site alternative ordered states by timestamp plus origin
+deployment ID. It would require a new origin field in madmin-go transport and
+a persisted origin field in `BucketMetadata`. That adds a dependency release,
+wire compatibility work, and an on-disk schema change without improving the
+convergence guarantee over the content-based total order.
+
+If a future product requirement needs provenance-aware conflict explanation,
+the source-site design can be introduced as a versioned protocol. It is not
+required to make the current register converge.
+
+## Bucket Lineage and `CreatedAt`
+
+`CreatedAt` protects a recreated bucket from delayed metadata events belonging
+to the prior bucket incarnation:
+
+```text
+if incoming.SourceUpdatedAt < local.CreatedAt:
+    ignore and log once per bucket
+```
+
+The floor is retained because removing it could install an old CORS grant on a
+new bucket with the same name. It is intentionally not used to classify the
+baseline.
+
+For current-version site replication, local events are generated strictly
+after `max(CreatedAt, current CORS barrier)`, and bucket creation timestamps are
+propagated before initial metadata sync. A floor rejection therefore indicates
+a stale lineage event, clock/history corruption, or a mixed/unsupported setup.
+The rejection is observable through a bucket-scoped log-once message and the
+remaining status mismatch.
+
+## Local Transition
+
+PUT and DELETE use the same CORS-specific transition helper.
+
+Under the bucket CORS namespace lock:
+
+1. load the current `.metadata.bin` through the migration-aware parsed loader;
+2. validate the new live payload, if any;
+3. choose:
+
+   ```text
+   UpdatedAt = max(UTCNow, CreatedAt + epsilon, CurrentBarrier + epsilon)
+   ```
+
+4. store either the live bytes or a nil tombstone with that timestamp;
+5. save and refresh the parsed cache; and
+6. release the lock before invoking `BucketMetaHook`.
+
+This preserves HTTP semantics while ensuring a local administrative action is
+strictly greater than the state it observed, including a future-dated peer
+barrier caused by clock skew. The peer path deliberately uses a raw metadata
+read instead: it must preserve the exact zero baseline and reject missing
+metadata rather than implicitly creating a peer bucket record.
+
+## Peer and Legacy-Bulk Transition
+
+Typed CORS dispatch decodes and validates the payload, then performs this join
+under the same lock:
+
+```text
+if incoming timestamp is zero:
+    reject
+if incoming timestamp is before CreatedAt:
+    ignore and log
+if incoming state <= local state:
+    no-op
+otherwise:
+    persist incoming payload and exact source timestamp
+```
+
+The admin handler's legacy/default bulk metadata path can also carry a non-nil
+CORS field. It therefore takes the same CORS lock, applies strict decoding and
+validation, and uses the same state comparison before saving. A nil CORS field
+in that untyped legacy shape means "not included" and cannot represent a
+tombstone; current producers use the typed CORS event for deletion.
+
+## Concurrency and Locking
+
+The transition lock is:
+
+```text
+.minio.sys / buckets/<bucket>/cors-config.lock
+```
+
+It is a virtual distributed namespace lock. The name deliberately differs from
+the real `buckets/<bucket>/.metadata.bin` object because the metadata save path
+locks that object internally and namespace locks are not re-entrant.
+
+The lock serializes every intentional current-version local, typed-peer,
+legacy-bulk, and local-heal CORS transition across nodes of one cluster. It
+cannot prevent an unrelated whole-record writer from restoring stale CORS
+columns. Residual paths include another metadata type's `Update`/`Delete`, a
+legacy bulk item whose nil CORS field means "not included",
+`ImportBucketMetadata`, and bucket-make metadata rewriting. Their inherited
+whole-record behavior is the separate architectural problem under issue #77.
+
+No cross-site admin call or `BucketMetaHook` dispatch is made while holding the
+CORS lock. The metadata save can perform blocking intra-cluster notification
+fan-out before the lock is released. Local handlers release the lock before
+cross-site dispatch; reordered network delivery is handled by the total-order
+join.
+
+## Dispatch and Retry
+
+PUT sends a typed `SRBucketMetaTypeCorsConfig` event with canonical base64 XML
+and the local source timestamp. DELETE sends the same type with `Cors == nil`
+and the tombstone timestamp.
+
+`BucketMetaHook` may deliver concurrently to sites, fail on a subset, or be
+retried by an external operation. The receiver transition is idempotent, so the
+transport does not need to impose a global event order.
+
+Current-version admin dispatch routes the typed event directly to
+`PeerBucketCorsConfigHandler`. The legacy/default path is hardened only to
+prevent a non-nil CORS field from bypassing the join; it is not a tombstone
+compatibility protocol.
+
+## Status Projection
+
+`SiteReplicationMetaInfo` always exports `CorsConfigUpdatedAt`, including zero
+baseline and nil tombstone states. It exports `CorsConfig` only for a live
+payload.
+
+Status considers sites converged if and only if every site has the same full
+CORS state:
+
+```text
+(kind, decoded payload bytes, SourceUpdatedAt)
+```
+
+Live payload counts remain useful for per-site summary totals, but they do not
+determine `CorsCfgMismatch`.
+
+Examples:
+
+| Site A | Site B | Mismatch |
+| --- | --- | --- |
+| baseline | baseline | no |
+| same live bytes at same timestamp | same live bytes at same timestamp | no |
+| same live bytes at different timestamps | same live bytes at different timestamps | yes |
+| same tombstone timestamp | same tombstone timestamp | no |
+| tombstones at different timestamps | tombstones at different timestamps | yes |
+| live | tombstone | yes |
+| invalid wire state | any state | yes |
+
+## Winner Selection and Heal
+
+Heal computes the maximum non-baseline valid state using the total order.
+Selection is independent of Go map iteration. Deployment ID is used only as a
+stable log-source choice when two sites already expose exactly equal states.
+
+For each different site:
+
+- the local site delegates to the normal peer CORS transition, preserving the
+  source timestamp and lock discipline;
+- a remote site receives a typed `SRBucketMetaTypeCorsConfig` event with the
+  winner's canonical payload or nil tombstone and exact timestamp.
+
+Payload equality alone is insufficient. A site with identical bytes at an
+older timestamp is healed so it acquires the same future ordering barrier.
+
+If every reported state is baseline, there is no event to propagate. If every
+reported state is invalid, status remains mismatched and heal does not select
+corrupt input as a source.
+
+## Initial Sync
+
+Initial sync emits:
+
+- a live event when `CorsConfigUpdatedAt` is non-zero and payload is live;
+- a tombstone event when `CorsConfigUpdatedAt` is non-zero and payload is nil;
+- no event for the zero baseline.
+
+Replaying initial sync is idempotent. A missed DELETE is recoverable because
+the tombstone is part of the snapshot rather than being inferred from the
+absence of a live payload.
+
+All sites must run the CORS-aware implementation before enabling or mutating
+per-bucket CORS. An older receiver can route an unknown typed event through a
+legacy path that cannot represent deletion and does not provide this ordering
+contract.
+
+## Persistence and Restart
+
+`CorsConfigXML` and `CorsConfigUpdatedAt` are persisted together in
+`BucketMetadata` msgpack. Zero time round-trips as zero; CORS is deliberately
+not defaulted to `CreatedAt` during load.
+
+`BucketMetadata.Save` parses the live CORS XML before writing and before the
+metadata system replaces the local cache. Therefore a rejected payload cannot
+poison disk or cache, and a successful peer/heal transition immediately serves
+the newly persisted parsed configuration.
+
+After cache removal or process restart:
+
+- a live state restores the same parsed rules and source timestamp;
+- a tombstone restores nil payload plus its non-zero timestamp;
+- a baseline remains nil plus zero timestamp.
+
+## Error Handling
+
+| Error | Behavior |
+| --- | --- |
+| zero source timestamp on a peer live/delete event | reject the event |
+| invalid/non-canonical base64 | reject before locking or saving |
+| empty non-nil payload | reject |
+| malformed XML | reject before saving |
+| semantically invalid CORS rules | reject before saving |
+| event before bucket `CreatedAt` | ignore and log once per bucket |
+| missing bucket metadata | return an error; do not create metadata implicitly |
+| exact duplicate or lower state | successful no-op |
+| remote heal failure | log the peer error; future heal cycles retry |
+
+## Alternatives Considered
+
+### Timestamp only
+
+Rejected. Ignoring or accepting every equal-timestamp conflict leaves an
+already divergent pair without a deterministic repair rule.
+
+### Timestamp plus source deployment ID
+
+Rejected for the current protocol. It is convergent, but requires madmin-go,
+wire, and persisted-schema changes without improving convergence over the
+selected total order.
+
+### Payload-only status and heal
+
+Rejected. It cannot distinguish ordering barriers and suppresses the exact
+heal needed to make later event acceptance consistent.
+
+### Default baseline timestamp to bucket creation
+
+Rejected. It conflates baseline classification with a mutable value that may
+come from a different snapshot and can turn a never-configured site into a
+false tombstone source.
+
+### Reuse `.metadata.bin` as the transition lock
+
+Rejected. The save path takes the same namespace lock internally; reusing it
+would self-deadlock.
+
+### Redesign every bucket metadata type together
+
+Rejected for issue #75. Neighboring metadata types have related inherited
+patterns but different delete, validation, and compatibility semantics. They
+require focused reproductions under issue #77.
+
+## Invariants
+
+The implementation is acceptable only while all of these invariants hold:
+
+1. Zero timestamp plus nil payload is the only baseline representation.
+2. Nil payload plus non-zero timestamp is a durable tombstone.
+3. A live payload has canonical base64 on the wire, valid CORS XML, and a
+   non-zero source timestamp.
+4. Every intentional current-version CORS state transition is serialized by
+   the CORS namespace lock from disk read through state comparison and save;
+   unrelated whole-record overwrite risk remains explicitly under issue #77.
+5. Peer apply and heal never replace local state with a lower or equal state.
+6. Local PUT and DELETE create a state strictly greater than the state observed
+   under the lock.
+7. A peer event before local bucket creation cannot modify the new bucket
+   lineage.
+8. Status reports convergence only for identical full states.
+9. Heal selects the same maximum regardless of arrival order, site, or map
+   iteration order.
+10. Same-payload/newer-timestamp heal advances the older barrier.
+11. Initial sync and retry preserve tombstones and source timestamps.
+12. Disk reload and cache reload preserve state kind, payload, and timestamp.
+
+## Test Contract
+
+The required test matrix is:
+
+| Area | Required evidence |
+| --- | --- |
+| Wire | canonical base64 accepted; case-different decoded bytes differ; malformed and non-canonical base64 rejected |
+| Validation | invalid XML and semantically invalid origin/method/rule rejected without mutation |
+| Ordering | older event ignored; newer event applied; duplicate no-op; equal live/live order-independent; equal PUT/DELETE chooses tombstone |
+| Barrier | same payload with newer timestamp is persisted and healed |
+| Tombstone | delayed PUT cannot resurrect; missed DELETE wins heal; repeated DELETE is idempotent |
+| Status | baseline, live, tombstone, payload mismatch, and timestamp-only mismatch classified correctly |
+| Winner | three-site equal-timestamp selection remains deterministic across repeated map iteration |
+| Concurrency | concurrent peer and legacy-bulk events converge to the total-order maximum |
+| Local concurrency | concurrent local PUT/DELETE timestamps are unique and final state matches the last serialized transition |
+| Initial sync | baseline omitted; live and tombstone emitted with exact source timestamp |
+| Lineage | pre-creation event ignored; post-creation event applied |
+| Restart | cache removal/disk reload preserves tombstone or live state and status timestamp |
+| Full seam | signed admin dispatch -> peer apply -> real status collection -> local heal -> cache reload -> remote heal dispatch |
+
+## Local Verification Record
+
+The current uncommitted implementation has passed:
+
+- the supplied adversarial base64 and same-payload/newer-timestamp tests;
+- focused CORS normal tests;
+- focused CORS race tests;
+- `go test ./internal/bucket/cors` and its race run;
+- `go test ./cmd -count=1`;
+- `go vet ./...`;
+- `go build ./...`;
+- repository-configured golangci-lint v2.13.1 with zero issues;
+- gofmt and `git diff --check`;
+- a signed admin dispatch -> apply -> status -> heal -> cache reload test.
+
+The repository `make lint` bootstrap could not download its private copy of
+golangci-lint because the network returned HTTP status 000. The same exact
+v2.13.1 binary already installed locally was used with the Makefile's build
+tags, timeout, and configuration and reported zero issues.
+
+## Independent Review Record
+
+Three read-only local Claude Code reviews used canonical model
+`claude-opus-5` at `max` effort.
+
+The first review rejected the pre-fix candidate and identified the unsafe
+CreatedAt-based baseline, missing deterministic tie-break, missing atomic join,
+non-monotonic local barrier, timestamp-blind status, and initial-sync tombstone
+gap. The selected C-prime model incorporated the valid findings while rejecting
+the suggestion to rewrite normal source timestamps.
+
+The second review found no P0. Its `GO WITH FIXES` findings were peer semantic
+validation, the legacy/default admin mutation path bypassing the CORS lock and
+join, CreatedAt-floor observability, and missing tests for invalid XML, lineage,
+and concurrent local transitions. Those required changes and tests are now in
+the working tree.
+
+The final review examined this design and the exact dirty diff, independently
+reran build, vet, lint, normal tests, and race tests, and found no P0 or P1.
+Its verdict was `GO WITH FIXES`: the implementation was explicitly judged GO,
+while five design-document statements required correction. It also suggested
+an optional status hardening so semantically invalid canonical payloads are
+not selected and retransmitted. The hardening and all mandatory documentation
+corrections are incorporated in the current tree. The final selected solution
+is therefore the C-prime register and invariants recorded in this document.
+
+## Release Gates
+
+An implementation-level GO means only that the local CORS state machine and
+tests satisfy this document. It does not authorize a release.
+
+Before closing issue #75 or publishing a server artifact:
+
+1. commit the exact reviewed implementation and design with DCO sign-off;
+2. push a focused branch and run remote PR CI;
+3. merge and confirm main CI on the merge commit;
+4. finish public EN/ZH upgrade, fallback, and downgrade documentation in
+   `silo.pgsty.com`;
+5. run a real two-site process test for PUT, DELETE, simultaneous conflict,
+   offline peer restart, status, and heal;
+6. verify no release tag or image contains an intermediate candidate; and
+7. treat package, image, SBOM, signature, canary, and production verification
+   as separate gates.

internal/bucket/cors/cors.go

@@ -87,6 +87,12 @@ func (c *Config) Validate() error {
 			return errors.New("CORSRule must contain at least one AllowedMethod")
 		}
 		for _, o := range r.AllowedOrigins {
+			if o == "" {
+				return errors.New("AllowedOrigin must not be empty")
+			}
+			if strings.Contains(o, "?") {
+				return errors.New("AllowedOrigin may not contain wildcard '?': " + o)
+			}
 			if strings.Count(o, "*") > 1 {
 				return errors.New("AllowedOrigin may contain at most one wildcard '*': " + o)
 			}
@@ -97,6 +103,9 @@ func (c *Config) Validate() error {
 			}
 		}
 		for _, h := range r.AllowedHeaders {
+			if strings.Contains(h, "?") {
+				return errors.New("AllowedHeader may not contain wildcard '?': " + h)
+			}
 			if strings.Count(h, "*") > 1 {
 				return errors.New("AllowedHeader may contain at most one wildcard '*': " + h)
 			}
@@ -108,14 +117,19 @@ func (c *Config) Validate() error {
 	return nil
 }
 
-// HasAllowedOrigin reports whether the rule allows the given origin.
+func (r Rule) matchAllowedOrigin(origin string) (string, bool) {
-func (r Rule) HasAllowedOrigin(origin string) bool {
 	for _, o := range r.AllowedOrigins {
 		if o == "*" || wildcard.MatchSimple(o, origin) {
-			return true
+			return o, true
 		}
 	}
-	return false
+	return "", false
+}
+
+// HasAllowedOrigin reports whether the rule allows the given origin.
+func (r Rule) HasAllowedOrigin(origin string) bool {
+	_, ok := r.matchAllowedOrigin(origin)
+	return ok
 }
 
 // HasAllowedMethod reports whether the rule allows the given HTTP method.
@@ -154,15 +168,17 @@ func (r Rule) headerAllowed(header string) bool {
 	return false
 }
 
-// MatchRule returns the first rule whose origin and method both match.
+// MatchRule returns the first rule whose origin and method both match, along
-func (c *Config) MatchRule(origin, method string) (*Rule, bool) {
+// with the configured origin pattern that matched.
+func (c *Config) MatchRule(origin, method string) (rule *Rule, allowedOrigin string, ok bool) {
 	for i := range c.CORSRules {
 		r := &c.CORSRules[i]
-		if r.HasAllowedOrigin(origin) && r.HasAllowedMethod(method) {
+		matchedOrigin, originOK := r.matchAllowedOrigin(origin)
-			return r, true
+		if originOK && r.HasAllowedMethod(method) {
+			return r, matchedOrigin, true
 		}
 	}
-	return nil, false
+	return nil, "", false
 }
 
 // MatchPreflight returns the first rule whose origin and method match and
@@ -170,17 +186,18 @@ func (c *Config) MatchRule(origin, method string) (*Rule, bool) {
 // this keeps evaluating subsequent rules until one fully satisfies the
 // preflight request, since an earlier origin/method match with a more
 // restrictive header list must not shadow a later, more permissive rule.
-func (c *Config) MatchPreflight(origin, method string, reqHeaders []string) (rule *Rule, allowedHeaders []string, ok bool) {
+func (c *Config) MatchPreflight(origin, method string, reqHeaders []string) (rule *Rule, allowedOrigin string, allowedHeaders []string, ok bool) {
 	for i := range c.CORSRules {
 		r := &c.CORSRules[i]
-		if !r.HasAllowedOrigin(origin) || !r.HasAllowedMethod(method) {
+		matchedOrigin, originOK := r.matchAllowedOrigin(origin)
+		if !originOK || !r.HasAllowedMethod(method) {
 			continue
 		}
 		allowed, headersOK := r.FilterAllowedHeaders(reqHeaders)
 		if !headersOK {
 			continue
 		}
-		return r, allowed, true
+		return r, matchedOrigin, allowed, true
 	}
-	return nil, nil, false
+	return nil, "", nil, false
 }

internal/bucket/cors/cors_test.go

@@ -55,10 +55,13 @@ func TestValidateRejections(t *testing.T) {
 	cases := map[string]string{
 		"bad method":            `<CORSConfiguration><CORSRule><AllowedOrigin>*</AllowedOrigin><AllowedMethod>TRACE</AllowedMethod></CORSRule></CORSConfiguration>`,
 		"no origin":             `<CORSConfiguration><CORSRule><AllowedMethod>GET</AllowedMethod></CORSRule></CORSConfiguration>`,
+		"empty origin":          `<CORSConfiguration><CORSRule><AllowedOrigin></AllowedOrigin><AllowedMethod>GET</AllowedMethod></CORSRule></CORSConfiguration>`,
 		"no method":             `<CORSConfiguration><CORSRule><AllowedOrigin>*</AllowedOrigin></CORSRule></CORSConfiguration>`,
 		"negative age":          `<CORSConfiguration><CORSRule><AllowedOrigin>*</AllowedOrigin><AllowedMethod>GET</AllowedMethod><MaxAgeSeconds>-1</MaxAgeSeconds></CORSRule></CORSConfiguration>`,
 		"multi wildcard origin": `<CORSConfiguration><CORSRule><AllowedOrigin>https://*.*.example.com</AllowedOrigin><AllowedMethod>GET</AllowedMethod></CORSRule></CORSConfiguration>`,
 		"multi wildcard header": `<CORSConfiguration><CORSRule><AllowedOrigin>*</AllowedOrigin><AllowedMethod>GET</AllowedMethod><AllowedHeader>x-*-*</AllowedHeader></CORSRule></CORSConfiguration>`,
+		"question mark origin":  `<CORSConfiguration><CORSRule><AllowedOrigin>https://?.example.com</AllowedOrigin><AllowedMethod>GET</AllowedMethod></CORSRule></CORSConfiguration>`,
+		"question mark header":  `<CORSConfiguration><CORSRule><AllowedOrigin>*</AllowedOrigin><AllowedMethod>GET</AllowedMethod><AllowedHeader>x-amz-?</AllowedHeader></CORSRule></CORSConfiguration>`,
 		"overlong id":           `<CORSConfiguration><CORSRule><ID>` + strings.Repeat("a", 256) + `</ID><AllowedOrigin>*</AllowedOrigin><AllowedMethod>GET</AllowedMethod></CORSRule></CORSConfiguration>`,
 	}
 	for name, doc := range cases {
@@ -74,14 +77,14 @@ func TestValidateRejections(t *testing.T) {
 
 func TestMatching(t *testing.T) {
 	c, _ := ParseBucketCorsConfig(strings.NewReader(sampleCORS))
-	rule, ok := c.MatchRule("https://api.example.org", "GET")
+	rule, _, ok := c.MatchRule("https://api.example.org", "GET")
 	if !ok {
 		t.Fatal("expected origin+method to match")
 	}
-	if _, ok := c.MatchRule("http://evil.com", "GET"); ok {
+	if _, _, ok := c.MatchRule("http://evil.com", "GET"); ok {
 		t.Fatal("did not expect match for disallowed origin")
 	}
-	if _, ok := c.MatchRule("http://www.example.com", "DELETE"); ok {
+	if _, _, ok := c.MatchRule("http://www.example.com", "DELETE"); ok {
 		t.Fatal("did not expect match for disallowed method")
 	}
 	allowed, ok := rule.FilterAllowedHeaders([]string{"x-amz-date", "x-amz-content-sha256"})
@@ -118,7 +121,7 @@ func TestMatchPreflightFallsThroughToLaterRule(t *testing.T) {
 		t.Fatalf("parse failed: %v", err)
 	}
 
-	rule, allowed, ok := c.MatchPreflight("https://app.example.com", "GET", []string{"x-custom-header"})
+	rule, _, allowed, ok := c.MatchPreflight("https://app.example.com", "GET", []string{"x-custom-header"})
 	if !ok {
 		t.Fatal("expected MatchPreflight to succeed via the later, permissive rule")
 	}
@@ -129,3 +132,26 @@ func TestMatchPreflightFallsThroughToLaterRule(t *testing.T) {
 		t.Fatalf("unexpected allowed headers: %v", allowed)
 	}
 }
+
+func TestMatchAllowedOriginReturnsFirstMatchingPattern(t *testing.T) {
+	rule := Rule{AllowedOrigins: []string{"https://app.example.com", "https://*", "*"}}
+
+	tests := []struct {
+		origin string
+		want   string
+	}{
+		{"https://app.example.com", "https://app.example.com"},
+		{"https://other.example.com", "https://*"},
+		{"http://other.example.com", "*"},
+	}
+
+	for _, tt := range tests {
+		got, ok := rule.matchAllowedOrigin(tt.origin)
+		if !ok {
+			t.Fatalf("expected %q to match", tt.origin)
+		}
+		if got != tt.want {
+			t.Fatalf("origin %q matched %q, want %q", tt.origin, got, tt.want)
+		}
+	}
+}