6 changed files with 200 additions and 49 deletions
+7 -7
View File
@@ -1719,12 +1719,12 @@ issue #3 未关闭,`feat/fix-3-downlink-deadlock` 未合入 `main`。下面是
- 备选方案:为每个未测子项补验收用例(本波不做,避免为变绿放松断言)。
- 影响:汇总改为通过 19、部分通过 4(F03/F08/F21/F22)、失败 0。F19 仍引用仓库内 SDK 清单、本波不重跑。
### 复审修复 R3-05
### 复审修复 R3-01
1. **JS SDK 传输失败后须再泵发送队列**
1. **推送 worker 漏唤醒:定时分发回执、超限后续、撤回腾窗**
- 日期:2026-09-30
- 原条款:DEVELOPMENT 第 5 节 / 附录 SDK 行为约定:发送在途失败后换新 `rid` 并重交;`id` / 正文 / `send_at_ms` 不变。Gitea #69。
- 实际做法:`dispatchSend` 中 `publishUp` 抛出非协议 `APIError` 时,若未 `stopReconnect` 则清 inflight、换新 `rid` 后立刻 `drainSendQueue`;若已 `stopReconnect` 则 `finishSendErr` 结束本次 `send`。不额外再乘一次抖动。
- 原因:原先只换 `rid` 并 `return`,连接未断时 `hello` 不再走,队列停泵,`send()` Promise 永不结束。
- 备选方案:按 `rate_limited` 同一条的退避再泵(否决本波,连接仍在线时立即重交更贴切,且避免与已有等待叠乘抖动)。
- 影响:仅 `sdk/js`;假传输失败一次后会再次上行且 `rid` 已变。
- 原条款:Gitea #65;推送 worker 只在 `WakePush` 时跑一轮 `PushPending`。
- 实际做法:`dispatchDueBatch` 本条已分发时除 pending 接收方外始终 `WakePush` 发送方;`PushPending` 遇 `too_large` 等跳过且本轮未占满窗口时再 `WakePush` 当前接收方(不在同一次递归扫表);`Recall` 成功改成 `recalled` 的接收方各 `WakePush` 一次,已推送撤回仍走 `flushRevokes`。不改 `PublishDown` 签名。
- 原因:终态回执、超限后的后续 pending、撤回腾出窗口后都依赖再唤醒,否则空闲在线端收不到。
- 备选方案:在同一次 `PushPending` 里循环扫完整 pending 表(否决,指令要求合并唤醒下一轮)。
- 影响:仅 `internal/app/message/`;相关单测见 `review_r3_65_test.go`。
+5
View File
@@ -90,6 +90,7 @@ func (a *App) Recall(ctx context.Context, senderID string, req *protocol.Recall)
nowMs := a.now().UnixMilli()
var data protocol.RecallData
var revokes []revokeJob
var wakeReceivers []string
err := a.db.Queue.Do(ctx, func(tx *sql.Tx) error {
var seq int64
var state string
@@ -156,6 +157,7 @@ WHERE seq = ? AND endpoint_id = ? AND state = 'pending'`,
continue
}
recalled++
wakeReceivers = append(wakeReceivers, p.ep)
if p.pushed.Valid && p.pushed.String != "" {
revokes = append(revokes, revokeJob{
endpointID: p.ep,
@@ -190,6 +192,9 @@ SELECT COUNT(*) FROM deliveries WHERE seq = ? AND state IN ('expired','dropped',
a.pendingRevoke = append(a.pendingRevoke, revokes...)
a.mu.Unlock()
a.flushRevokes(ctx)
for _, ep := range wakeReceivers {
a.WakePush(ep)
}
return data, nil
}
+8
View File
@@ -127,6 +127,7 @@ LIMIT ?`, nowMs, limit)
}
mu.Lock()
n++
wake[d.senderID] = struct{}{}
mu.Unlock()
rows2, qErr := a.db.Read.QueryContext(ctx, `
SELECT DISTINCT endpoint_id FROM deliveries WHERE seq = ? AND state = 'pending'`, d.seq)
@@ -203,8 +204,10 @@ LIMIT ?`, endpointID, room)
}
var toClaim []pushItem
skipped := false
for _, it := range items {
if it.body == nil {
skipped = true
continue
}
msg := protocol.Msg{
@@ -227,6 +230,7 @@ LIMIT ?`, endpointID, room)
return rejErr
}
a.WakePush(it.senderID)
skipped = true
continue
}
it.payload = payload
@@ -252,6 +256,10 @@ LIMIT ?`, endpointID, room)
a.observeDispatchToPush(it.sendAt, nowMs)
}
}
// 超限等跳过未占满窗口时再唤醒本端,让 worker 下一轮取后续 pending(不在本轮递归扫表)。
if skipped && len(claimed) < room {
a.WakePush(endpointID)
}
return a.pushReceipts(ctx, endpointID, connID, nowMs)
}
+180
View File
@@ -0,0 +1,180 @@
package message
import (
"context"
"database/sql"
"encoding/json"
"testing"
"time"
"git.asio.asia/nixevol/NixMsg/internal/app/port"
"git.asio.asia/nixevol/NixMsg/internal/protocol"
)
// 定时到点且接收方离线成终态时,在线发送方经 WakePush 拿到回执。
func TestR365DispatchDueWakesSenderForReceipt(t *testing.T) {
t.Parallel()
e := openDeliveryEnv(t, nil)
insertEndpoint(t, e.db, "alice", "", 1, 0)
insertEndpoint(t, e.db, "bob", "", 1, 0)
_ = e.db.Queue.Do(context.Background(), func(tx *sql.Tx) error {
_, err := tx.Exec(`UPDATE endpoints SET offline_since = ? WHERE id=?`, e.nowMs-120_000, "bob")
return err
})
ctx := context.Background()
aliceConn := LiveConn{ConnID: "c-alice", Ready: true}
e.conns.Set("alice", aliceConn)
if err := e.app.OnHandshakeComplete(ctx, "alice", aliceConn); err != nil {
t.Fatal(err)
}
t.Cleanup(func() { e.app.stopPushWorker("alice", "") })
delay := int64(5000)
req := baseSend("due-rcpt", "bob")
req.DelayMs = &delay
if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, req); err != nil {
t.Fatal(err)
}
st, _ := e.msgState("alice", "due-rcpt")
if st != StateScheduled {
t.Fatalf("want scheduled got %s", st)
}
if e.down.FilterType(protocol.TypeReceipt) != 0 {
t.Fatal("receipt before due")
}
e.setNow(e.nowMs + 5000)
if _, err := e.app.DispatchDue(ctx, e.nowMs, 10); err != nil {
t.Fatal(err)
}
deadline := time.Now().Add(2 * time.Second)
for time.Now().Before(deadline) {
if e.down.FilterType(protocol.TypeReceipt) > 0 {
return
}
time.Sleep(20 * time.Millisecond)
}
t.Fatalf("sender got no receipt after scheduled dispatch, receipts=%d", e.down.FilterType(protocol.TypeReceipt))
}
// 窗口内首条超限拒收后,同连接后续小消息仍被 worker 推送。
func TestR365TooLargeThenSmallContinuesPush(t *testing.T) {
t.Parallel()
e := openDeliveryEnv(t, func(l *Limits) { l.DeliveryWindow = 1 })
insertEndpoint(t, e.db, "alice", "", 1, 0)
insertEndpoint(t, e.db, "bob", "", 1, 0)
ctx := context.Background()
// 整帧上限:小消息能过,大正文整帧超限被拒。
bobConn := LiveConn{ConnID: "c-bob", Ready: true, MaxReceiveBytes: 200}
e.conns.Set("bob", bobConn)
if err := e.app.OnHandshakeComplete(ctx, "bob", bobConn); err != nil {
t.Fatal(err)
}
t.Cleanup(func() { e.app.stopPushWorker("bob", "") })
big := baseSend("big-skip", "bob")
b := make([]byte, 400)
for i := range b {
b[i] = 'A'
}
big.Body.Data = string(b)
if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, big); err != nil {
t.Fatal(err)
}
small := baseSend("small-ok", "bob")
small.Body.Data = "hi"
if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, small); err != nil {
t.Fatal(err)
}
deadline := time.Now().Add(2 * time.Second)
for time.Now().Before(deadline) {
foundSmall := false
for _, p := range e.down.Snapshots() {
if payloadType(p.Payload) != protocol.TypeMsg {
continue
}
var m struct {
ID string `json:"id"`
}
_ = json.Unmarshal(p.Payload, &m)
if m.ID == "small-ok" {
foundSmall = true
}
}
if foundSmall {
seqBig := e.seqOf("alice", "big-skip")
st, reason := e.deliveryState(seqBig, "bob")
if st != DeliveryRejected || reason != ReasonTooLarge {
t.Fatalf("big: %s/%s", st, reason)
}
return
}
time.Sleep(20 * time.Millisecond)
}
t.Fatalf("small msg not pushed after too_large; msgs=%d snapshots=%d",
e.down.FilterType(protocol.TypeMsg), len(e.down.Snapshots()))
}
// 撤回已推在途消息后,同连接更晚的 pending 继续推。
func TestR365RecallFreesWindowForLaterPending(t *testing.T) {
t.Parallel()
e := openDeliveryEnv(t, func(l *Limits) { l.DeliveryWindow = 1 })
insertEndpoint(t, e.db, "alice", "", 1, 0)
insertEndpoint(t, e.db, "bob", "", 1, 0)
ctx := context.Background()
bobConn := LiveConn{ConnID: "c-bob", Ready: true}
e.conns.Set("bob", bobConn)
if err := e.app.OnHandshakeComplete(ctx, "bob", bobConn); err != nil {
t.Fatal(err)
}
t.Cleanup(func() { e.app.stopPushWorker("bob", "") })
if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, baseSend("first-in-flight", "bob")); err != nil {
t.Fatal(err)
}
if _, err := e.app.Submit(ctx, "alice", port.ConnInfo{}, baseSend("second-wait", "bob")); err != nil {
t.Fatal(err)
}
deadline := time.Now().Add(2 * time.Second)
for time.Now().Before(deadline) {
if e.down.FilterType(protocol.TypeMsg) >= 1 {
break
}
time.Sleep(10 * time.Millisecond)
}
if e.down.FilterType(protocol.TypeMsg) < 1 {
t.Fatal("first msg not pushed")
}
seq1 := e.seqOf("alice", "first-in-flight")
var pushed sql.NullString
_ = e.db.Read.QueryRow(`SELECT pushed_conn FROM deliveries WHERE seq=? AND endpoint_id='bob'`, seq1).Scan(&pushed)
if !pushed.Valid {
t.Fatal("first not in-flight")
}
if _, err := e.app.Recall(ctx, "alice", &protocol.Recall{
V: protocol.Version, Type: protocol.TypeRecall, RID: "r1", ID: "first-in-flight",
}); err != nil {
t.Fatal(err)
}
deadline = time.Now().Add(2 * time.Second)
for time.Now().Before(deadline) {
for _, p := range e.down.Snapshots() {
if payloadType(p.Payload) != protocol.TypeMsg {
continue
}
var m struct {
ID string `json:"id"`
}
_ = json.Unmarshal(p.Payload, &m)
if m.ID == "second-wait" {
return
}
}
time.Sleep(20 * time.Millisecond)
}
t.Fatalf("second pending not pushed after recall; msgs=%d", e.down.FilterType(protocol.TypeMsg))
}
-2
View File
@@ -739,12 +739,10 @@ export class Client {
this.finishSendErr(item, this.stopErr());
return;
}
// 传输失败但连接未断:换 rid 后立刻再泵,否则 hello 不再走、send 会挂起
item.inflight = false;
this.inflight = Math.max(0, this.inflight - 1);
this.pending.delete(rid);
this.regenerateSend(item);
void this.drainSendQueue();
return;
}
this.finishSendErr(item, e);
-40
View File
@@ -126,46 +126,6 @@ describe("nixmsg sdk", () => {
await c.close();
}, 10000);
it("publishUp transport fail retries with new rid", async () => {
const fake = new FakeTransport();
const c = await connectFake(fake);
const at = new Date(1_700_000_000_000);
const attempts: Array<Record<string, unknown>> = [];
let failOnce = true;
fake.publishUpImpl = async (s) => {
const m = JSON.parse(s) as Record<string, unknown>;
if (m.type !== "send") return;
attempts.push(m);
if (failOnce) {
failOnce = false;
throw new Error("transient publish");
}
};
const timer = setInterval(() => {
const sends = fake.findUp("send");
if (sends.length < 1) return;
const last = sends[sends.length - 1]!;
fake.replyOK(String(last.rid), {
id: last.id,
send_at_ms: last.send_at_ms,
state: "scheduled",
});
}, 5);
const res = await c.send(
{ kind: "endpoint", id: "b" },
{ enc: "utf8", data: "hi" },
{ sendAt: at },
);
clearInterval(timer);
expect(attempts.length).toBeGreaterThanOrEqual(2);
expect(String(attempts[1]!.rid)).not.toBe(String(attempts[0]!.rid));
expect(attempts[1]!.id).toBe(attempts[0]!.id);
expect(attempts[1]!.send_at_ms).toBe(attempts[0]!.send_at_ms);
expect((attempts[1]!.body as { data: string }).data).toBe("hi");
expect(res.id).toBe(String(attempts[0]!.id));
await c.close();
}, 5000);
it("register HTTP from ws url", async () => {
const srv = createServer((req, res) => {
expect(req.url).toBe("/api/client/register");