fix(images): 服务器镜像在创建时固定到仓库 digest，更换镜像需确认备份，安装器重建前先固定旧服并推送不可变版本标签

internal/api/api.go

@@ -33,6 +33,11 @@ type API struct {
 	// still exercised even before the subsystem is wired in.
 	Builder ImageBuilder
 
+	// Images pins a whitelisted image ref to the digest it names when a server is
+	// created or its image is changed (internal/imagepin), so a later push over
+	// the same tag never reaches an existing world. Nil stores refs as given.
+	Images ImagePinner
+
 	// Console is the synchronous RCON write channel (spec §8 写=RCON). It is
 	// wired in production (cmd/felis); a nil Console makes the command route report
 	// 503 rather than panic, so the ownership boundary is still exercised in tests.

internal/api/audit.go

@@ -64,6 +64,19 @@ func (a *API) audit(r *http.Request, action, target string) {
 	a.auditEntry(r, e)
 }
 
+// auditImageChange records a confirmed image change as server.patch with the
+// image it replaced and the one it set, so the audit log alone can say which
+// build a world ran before it was moved.
+func (a *API) auditImageChange(r *http.Request, server, from, to string) {
+	p := principalFromContext(r.Context())
+	e := AuditEntry{Actor: auditActor(p), Action: "server.patch", ServerName: server}
+	if p != nil {
+		e.ActorUserID = p.UserID
+	}
+	e.Payload = auditPayload(map[string]any{"image_from": from, "image_to": to})
+	a.auditEntry(r, e)
+}
+
 // auditAccount records an action a pre-session door took for the account it
 // resolved (u nil: none was). The username is the actor: the door has not yet
 // proven anything about the address.

internal/api/handlers_patch_test.go

@@ -75,7 +75,7 @@ func TestPatchServerAutostartPolicy(t *testing.T) {
 func TestPatchServerImageReAdmitted(t *testing.T) {
 	api, _, cl, _ := newPatchAPI()
 
-	w := patchSurvival(api, `{"image":"`+admittedImage+`"}`)
+	w := patchSurvival(api, `{"image":"`+admittedImage+`","confirmImageChange":true}`)
 	if w.Code != http.StatusOK {
 		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
 	}

internal/api/handlers_user.go

@@ -375,6 +375,13 @@ func (a *API) handleCreateServer(w http.ResponseWriter, r *http.Request) {
 			"image %q is not on the whitelist", body.Image))
 		return
 	}
+	// The spec keeps the digest the tag names now, not the tag: the world is
+	// created on this build and stays on it until an admin changes the image.
+	image, err := a.pinImage(r.Context(), body.Image)
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
 
 	// Quota is intentionally NOT enforced here. §15 creates an UNOWNED server
 	// (owner_id NULL); the per-user quota is charged at claim time (spec §9.3 /
@@ -432,7 +439,7 @@ func (a *API) handleCreateServer(w http.ResponseWriter, r *http.Request) {
 		Name:            body.Name,
 		Subdomain:       body.Subdomain,
 		DisplayName:     body.DisplayName,
-		Image:           body.Image,
+		Image:           image,
 		JavaMemory:      javaMemory,
 		StorageSize:     storage,
 		AutostartPolicy: policy,
@@ -613,11 +620,16 @@ func parsePositiveQuantity(s, field string) (resource.Quantity, error) {
 // the dual-write routing identity (name is the immutable object key; subdomain
 // would desync the Postgres alias) nor for the world PVC size (see below).
 type patchServerRequest struct {
-	DisplayName     *string          `json:"displayName,omitempty"`
-	AutostartPolicy *string          `json:"autostartPolicy,omitempty"`
-	Image           *string          `json:"image,omitempty"`
-	Memory          *string          `json:"memory,omitempty"`
-	Resources       *resourceRequest `json:"resources,omitempty"`
+	DisplayName     *string `json:"displayName,omitempty"`
+	AutostartPolicy *string `json:"autostartPolicy,omitempty"`
+	Image           *string `json:"image,omitempty"`
+	// ConfirmImageChange acknowledges that a new image opens the world with
+	// whatever Minecraft version it carries. Chunks a newer version has upgraded
+	// cannot be read by the older one again, so without it an image change that
+	// would actually move the server is refused (image_change_unconfirmed).
+	ConfirmImageChange bool             `json:"confirmImageChange,omitempty"`
+	Memory             *string          `json:"memory,omitempty"`
+	Resources          *resourceRequest `json:"resources,omitempty"`
 	// IdleStopSeconds sets idle auto-stop: 0 turns it off, otherwise the server
 	// stops after that many seconds with nobody online (60 to 86400).
 	IdleStopSeconds *int32 `json:"idleStopSeconds,omitempty"`
@@ -672,6 +684,8 @@ func (a *API) handlePatchServer(w http.ResponseWriter, r *http.Request) {
 	// so the response and audit name the real mutation.
 	var patch ServerSpecPatch
 	var changed []string
+	// imageFrom is the image a confirmed image change replaced, for the audit row.
+	var imageFrom string
 
 	if body.DisplayName != nil {
 		patch.DisplayName = body.DisplayName
@@ -730,8 +744,31 @@ func (a *API) handlePatchServer(w http.ResponseWriter, r *http.Request) {
 				"image %q is not on the whitelist", *body.Image))
 			return
 		}
-		patch.Image = body.Image
-		changed = append(changed, "image")
+		image, err := a.pinImage(r.Context(), *body.Image)
+		if err != nil {
+			writeError(w, r, err)
+			return
+		}
+		info, err := a.Cluster.GetServer(r.Context(), name)
+		if err != nil {
+			a.writeLookupError(w, r, err)
+			return
+		}
+		// Re-picking the tag a server was created from resolves to that tag's
+		// newest build, which is as much a version move as picking another image.
+		// Only a pin that lands on exactly the current image is no change at all.
+		if image != info.Image {
+			if !body.ConfirmImageChange {
+				writeError(w, r, newError(http.StatusConflict, "image_change_unconfirmed",
+					"changing the image from %q to %q opens this world with the new image's Minecraft version, "+
+						"and chunks it upgrades cannot be opened by the old one again; back the world up first, "+
+						"then resend with confirmImageChange", info.Image, image))
+				return
+			}
+			patch.Image = &image
+			changed = append(changed, "image")
+			imageFrom = info.Image
+		}
 	}
 
 	// Memory and the resource overrides move together: resolveResources derives the
@@ -814,7 +851,11 @@ func (a *API) handlePatchServer(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
-	a.audit(r, "server.patch", name)
+	if patch.Image != nil {
+		a.auditImageChange(r, name, imageFrom, *patch.Image)
+	} else {
+		a.audit(r, "server.patch", name)
+	}
 	writeJSON(w, http.StatusOK, map[string]any{
 		"name":    name,
 		"patched": changed,

internal/api/images.go

@@ -6,6 +6,7 @@ import (
 	"net/http"
 
 	"felis.lolicon.best/internal/build"
+	"felis.lolicon.best/internal/imagepin"
 	"k8s.io/apimachinery/pkg/util/validation"
 )
 
@@ -252,3 +253,29 @@ func writeBuildError(w http.ResponseWriter, r *http.Request, err error) {
 		writeError(w, r, err)
 	}
 }
+
+// ImagePinner resolves an image ref to the immutable form a server's spec keeps
+// (imagepin.Resolver). A ref it does not manage comes back unchanged.
+type ImagePinner interface {
+	Pin(ctx context.Context, ref string) (string, error)
+}
+
+// pinImage pins an admitted ref for a server spec. A tag the registry does not
+// hold is the caller's to fix (build or push it first); any other failure is the
+// registry being unreachable, and the server is not created or changed without a
+// pin, since an unpinned ref is exactly what lets a later push move its world.
+func (a *API) pinImage(ctx context.Context, ref string) (string, error) {
+	if a.Images == nil {
+		return ref, nil
+	}
+	pinned, err := a.Images.Pin(ctx, ref)
+	switch {
+	case errors.Is(err, imagepin.ErrNotFound):
+		return "", newError(http.StatusBadRequest, "image_not_in_registry",
+			"image %q is whitelisted but the registry does not hold it; build or push it first", ref)
+	case err != nil:
+		return "", newError(http.StatusServiceUnavailable, "registry_unavailable",
+			"could not resolve image %q to a digest: %v", ref, err)
+	}
+	return pinned, nil
+}

internal/api/images_pin_test.go

@@ -0,0 +1,162 @@
+package api
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"testing"
+
+	"felis.lolicon.best/internal/imagepin"
+)
+
+const pinnedDigest = "sha256:1111111111111111111111111111111111111111111111111111111111111111"
+
+// fakePinner pins every unpinned ref to pinnedDigest, or fails with err. A ref
+// that already names a digest comes back as is, like imagepin.Resolver.
+type fakePinner struct {
+	err  error
+	seen []string
+}
+
+func (f *fakePinner) Pin(_ context.Context, ref string) (string, error) {
+	f.seen = append(f.seen, ref)
+	if f.err != nil {
+		return "", f.err
+	}
+	if imagepin.Pinned(ref) {
+		return ref, nil
+	}
+	return ref + "@" + pinnedDigest, nil
+}
+
+// TestCreateServerStoresPinnedImage: the spec a server is created with names the
+// digest its tag resolved to, so a later push over the tag cannot move it.
+func TestCreateServerStoresPinnedImage(t *testing.T) {
+	api, _, cl, _ := newCreateAPI()
+	pin := &fakePinner{}
+	api.Images = pin
+
+	w := do(api.ExternalHandler(), "POST", "/api/v1/servers", validCreateBody, nil)
+	if w.Code != http.StatusCreated {
+		t.Fatalf("code = %d, want 201 (%s)", w.Code, w.Body.String())
+	}
+	if got, want := cl.created["survival"].Image, admittedImage+"@"+pinnedDigest; got != want {
+		t.Errorf("created image = %q, want %q", got, want)
+	}
+	if len(pin.seen) != 1 || pin.seen[0] != admittedImage {
+		t.Errorf("pinned refs = %v, want the admitted ref once", pin.seen)
+	}
+}
+
+// TestCreateServerPinErrors: a tag the registry does not hold is the caller's
+// problem (400); a registry that cannot answer is the platform's (503). Neither
+// creates anything.
+func TestCreateServerPinErrors(t *testing.T) {
+	for _, tc := range []struct {
+		err      error
+		code     int
+		wantCode string
+	}{
+		{fmt.Errorf("resolve: %w", imagepin.ErrNotFound), http.StatusBadRequest, "image_not_in_registry"},
+		{errors.New("dial tcp: connection refused"), http.StatusServiceUnavailable, "registry_unavailable"},
+	} {
+		api, _, cl, _ := newCreateAPI()
+		api.Images = &fakePinner{err: tc.err}
+
+		w := do(api.ExternalHandler(), "POST", "/api/v1/servers", validCreateBody, nil)
+		if w.Code != tc.code || decodeErr(t, w) != tc.wantCode {
+			t.Errorf("%v: got %d %s, want %d %s", tc.err, w.Code, w.Body.String(), tc.code, tc.wantCode)
+		}
+		if _, ok := cl.created["survival"]; ok {
+			t.Errorf("%v: server created despite the pin failure", tc.err)
+		}
+	}
+}
+
+// TestPatchServerImageChangeUnconfirmed: moving a world to another build is
+// refused until the caller acknowledges the chunk upgrade cannot be undone.
+func TestPatchServerImageChangeUnconfirmed(t *testing.T) {
+	api, repo, cl, _ := newPatchAPI()
+	cl.byName["survival"].Image = "registry.felis.svc:5000/mc:0@" + pinnedDigest
+	api.Images = &fakePinner{}
+
+	w := patchSurvival(api, `{"image":"`+admittedImage+`"}`)
+	if w.Code != http.StatusConflict || decodeErr(t, w) != "image_change_unconfirmed" {
+		t.Fatalf("got %d %s, want 409 image_change_unconfirmed", w.Code, w.Body.String())
+	}
+	if _, ok := cl.patched["survival"]; ok {
+		t.Error("spec patched without confirmation")
+	}
+	if len(repo.audits) != 0 {
+		t.Errorf("refused change audited: %+v", repo.audits)
+	}
+}
+
+// TestPatchServerImageConfirmedAudited: a confirmed change stores the pinned ref
+// and the audit row names both builds.
+func TestPatchServerImageConfirmedAudited(t *testing.T) {
+	api, repo, cl, _ := newPatchAPI()
+	from := "registry.felis.svc:5000/mc:0@" + pinnedDigest
+	cl.byName["survival"].Image = from
+	api.Images = &fakePinner{}
+
+	w := patchSurvival(api, `{"image":"`+admittedImage+`","confirmImageChange":true}`)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	to := admittedImage + "@" + pinnedDigest
+	if p := cl.patched["survival"]; p.Image == nil || *p.Image != to {
+		t.Fatalf("patched image = %v, want %q", p.Image, to)
+	}
+	if len(repo.audits) != 1 || repo.audits[0].Action != "server.patch" {
+		t.Fatalf("audits = %+v, want one server.patch", repo.audits)
+	}
+	var payload map[string]string
+	if err := json.Unmarshal(repo.audits[0].Payload, &payload); err != nil {
+		t.Fatalf("audit payload %q: %v", repo.audits[0].Payload, err)
+	}
+	if payload["image_from"] != from || payload["image_to"] != to {
+		t.Errorf("audit payload = %v, want image_from %q image_to %q", payload, from, to)
+	}
+}
+
+// TestPatchServerImageSamePinIsNoChange: re-picking the tag a server runs, while
+// the tag still names the same build, changes nothing and needs no confirmation.
+func TestPatchServerImageSamePinIsNoChange(t *testing.T) {
+	api, repo, cl, _ := newPatchAPI()
+	cl.byName["survival"].Image = admittedImage + "@" + pinnedDigest
+	api.Images = &fakePinner{}
+
+	w := patchSurvival(api, `{"image":"`+admittedImage+`"}`)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	if p := cl.patched["survival"]; p.Image != nil {
+		t.Errorf("patched image = %q, want no image change", *p.Image)
+	}
+	if len(repo.audits) != 1 || repo.audits[0].Payload != nil {
+		t.Errorf("audits = %+v, want a plain server.patch", repo.audits)
+	}
+}
+
+// TestPatchServerImageRestoresPinnedBuild: the exact build an earlier change
+// replaced is admitted by its tag and set as is, so a world restored from a
+// backup can go back to the build that wrote it.
+func TestPatchServerImageRestoresPinnedBuild(t *testing.T) {
+	api, _, cl, fb := newPatchAPI()
+	cl.byName["survival"].Image = admittedImage + "@" + pinnedDigest
+	pin := &fakePinner{}
+	api.Images = pin
+	old := admittedImage + "@sha256:" + fmt.Sprintf("%064d", 0)
+	fb.admitted[old] = true
+
+	w := patchSurvival(api, `{"image":"`+old+`","confirmImageChange":true}`)
+	if w.Code != http.StatusOK {
+		t.Fatalf("code = %d, want 200 (%s)", w.Code, w.Body.String())
+	}
+	if p := cl.patched["survival"]; p.Image == nil || *p.Image != old {
+		t.Errorf("patched image = %v, want %q", p.Image, old)
+	}
+}