fix(relay): 每个节点的会话与挑战数量设上限;agent 侧加 fuzz

设计假定节点是敌对方("被攻陷的节点不该能读到另一个节点的流量"),而中继为它
保存两份状态:会话(TTL 30 分钟)与挑战(TTL 60 秒)。两者此前只在**过期**时被
清理,所以一个循环握手的节点可以让它们无界增长——实测 5000ms 中继时间内 5000 次
握手留下 5000 个存活会话,且没有任何东西阻止接下来的 5000 次。

诚实值是 1:每节点只有一个连接,注册表也只保留一个挂起的轮询。所以上限取
每节点 4 个会话、16 个挑战,超出时淘汰最旧的而非拒绝——"最新一次握手获胜"已经是
注册表对连接本身的规则,拒绝会让一次合法的重连被自己早先的尝试挡住。

写这个上限时踩到并修掉了一个自己造的 bug:第一版保留最旧、删掉刚加入的,于是
循环握手的节点会被自己下一次握手淘汰掉最新会话,每次轮询都 401。只断言"表很小"
的测试抓不到它——能抓住它的是"最新会话仍然可用"这条断言。

agent 侧新增 fuzz(agent/internal/agent/fuzz_test.go):Handle 是唯一一个输入来自
被假定敌对一方的函数,而 agent 是单进程,这里的一次 panic 不是一次坏请求而是节点
下线直到 systemd 拉起。断言故意弱——不 panic,且要么返回错误要么返回合法 JSON——
拒绝逻辑由旁边的表驱动测试覆盖。跑 45 秒、134 万次执行,无失败;测试内跑的是种子
语料(Go 的常规做法,fuzztime 只在显式 -fuzz 时生效)。

Node 112 项、Go 全包、Swift 101 项通过。
This commit is contained in:
chick
2026-09-17 01:29:55 +08:00
parent e2d4ea317c
commit 72b44398cd
2 changed files with 55 additions and 3 deletions
+3 -3
View File
@@ -26,7 +26,7 @@ type pair struct {
appKx []byte
}
func newPair(t *testing.T) pair {
func newPair(t testing.TB) pair {
t.Helper()
_, sign, err := ed25519.GenerateKey(rand.Reader)
if err != nil {
@@ -51,7 +51,7 @@ type harness struct {
sentEphemeral map[string][]byte
}
func newHarness(t *testing.T, handler http.Handler) *harness {
func newHarness(t testing.TB, handler http.Handler) *harness {
t.Helper()
server := httptest.NewServer(handler)
t.Cleanup(server.Close)
@@ -105,7 +105,7 @@ func newHarness(t *testing.T, handler http.Handler) *harness {
}
// seal builds a complete request envelope the way the phone does.
func (h *harness) seal(t *testing.T, reqID string, inner wire.InnerRequest) []byte {
func (h *harness) seal(t testing.TB, reqID string, inner wire.InnerRequest) []byte {
t.Helper()
body, err := json.Marshal(inner)
if err != nil {
+52
View File
@@ -0,0 +1,52 @@
package agent
import (
"context"
"encoding/json"
"net/http"
"testing"
"time"
"fun.chick/msa-agent/internal/wire"
)
// Fuzzing the agent's entry point, because it is the one function whose input comes from a party the
// design assumes is hostile.
//
// The relay is untrusted, so every byte `Handle` sees was chosen by something that may be trying to
// take the node down rather than to command it. A panic in here is not a bad request: the agent is a
// single process holding one poll, so an unrecovered panic ends it, and the node goes offline until
// systemd restarts it. That makes "does not panic on arbitrary input" a security property rather
// than a robustness nicety.
//
// The assertion is deliberately weak — no panic, and either a response or an error — because the
// interesting behaviour (refusals, ordering) is covered by the table tests next door. What this
// covers is the space those tests cannot enumerate: malformed base64, truncated JSON, absurd
// lengths, and the combinations of valid-looking fields that no hand-written case would think of.
func FuzzHandleDoesNotPanicOnArbitraryEnvelopes(f *testing.F) {
// Seeds: a valid envelope, one with each field corrupted, and a few degenerate shapes. Fuzzing
// from nothing spends its first thousand iterations rediscovering that `{}` is not JSON.
h := newHarness(f, http.HandlerFunc(func(http.ResponseWriter, *http.Request) {}))
valid := h.seal(f, "3f2504e0-4f89-41d3-9a0c-0305e82c3301", wire.InnerRequest{Method: "GET", Path: "/api/health"})
f.Add(valid)
f.Add([]byte(`{}`))
f.Add([]byte(`[]`))
f.Add([]byte(`null`))
f.Add([]byte(`{"v":1,"node":"nd_test0001","req":"x","app":"y","ts":0,"ek":"","alg":"","p":"","s":""}`))
f.Add([]byte(`{"v":1,"node":"nd_test0001","req":"3f2504e0-4f89-41d3-9a0c-0305e82c3301","app":"ap_test0000","ts":18446744073709551615,"ek":"AQID","alg":"X25519-ED25519-AES256GCM-HKDFSHA256","p":"AAEC/v8=","s":"c2ln"}`))
f.Fuzz(func(t *testing.T, raw []byte) {
// A fresh harness per iteration would be far too slow, and the harness is read-only with
// respect to the agent for a refused request. The context is bounded so a handler that
// somehow blocks cannot stall the fuzzer.
ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
defer cancel()
response, err := h.agent.Handle(ctx, raw)
// The contract: an envelope the agent will not act on either returns an error (it was not
// for this node, or was unparseable) or returns a response that is itself valid JSON. A
// third outcome — a panic, or a response that is not JSON — is the bug.
if err == nil && len(response) > 0 && !json.Valid(response) {
t.Fatalf("Handle returned a response that is not valid JSON: %q", response)
}
})
}