fix: 补三处推送漏唤醒让回执与后续 pending 继续推 (#65 )

2026-09-30 17:29:09 +08:00
7 changed files with 212 additions and 104 deletions
@@ -1719,11 +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 R3-04
+### 复审修复 R3-01

- 日期：2026-09-30
- 原条款：Gitea #68；`dispatchSend` 在 `PublishUp` 失败且未停止重连时清 inflight 后静默 return，`Send` 永久挂起。
- 实际做法：失败且未 `stopReconnect` 时保留 id/body/send_at_ms，经 `regenerateSendLocked` 换新 rid，清本次 inflight 并 `drainSendQueue` 继续泵；已停止重连时仍 `finishSend` 返回错误。
- 原因：条目已不在途，重连时的 `requeueInflightLocked` 不会捡回，调用方一直等 `result`。
- 备选方案：失败即 `finishSend` 报错（否决，与断线/限速重交语义不一致）；只换 rid 不唤醒队列（否决，等同 JS #69）。
- 影响：仅 `sdk/go`；假传输可模拟 send 发布失败一次。
+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：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`。
@@ -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
 }

@@ -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)
 }

@@ -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))
+}
@@ -156,66 +156,6 @@ func TestFrameTooLargeLocal(t *testing.T) {
 	}
 }

-func TestPublishUpFailRetriesWithNewRID(t *testing.T) {
-	fake := NewFakeTransport()
-	fake.FailSendPublishN(1)
-	c := connectFake(t, fake)
-	defer c.Close()
-
-	at := time.UnixMilli(1_700_000_000_000)
-	fixedID := "msg-publish-retry"
-	done := make(chan struct{})
-	var firstRID, secondRID string
-
-	go func() {
-		defer close(done)
-		deadline := time.Now().Add(3 * time.Second)
-		for time.Now().Before(deadline) {
-			sends := fake.FindUp("send")
-			if len(sends) < 2 {
-				time.Sleep(5 * time.Millisecond)
-				continue
-			}
-			first, second := sends[0], sends[1]
-			firstRID, _ = first["rid"].(string)
-			secondRID, _ = second["rid"].(string)
-			if first["id"] != fixedID || second["id"] != fixedID {
-				t.Errorf("id changed: %v -> %v", first["id"], second["id"])
-			}
-			if first["send_at_ms"] != second["send_at_ms"] {
-				t.Errorf("send_at_ms changed: %v -> %v", first["send_at_ms"], second["send_at_ms"])
-			}
-			body1, _ := json.Marshal(first["body"])
-			body2, _ := json.Marshal(second["body"])
-			if string(body1) != string(body2) {
-				t.Errorf("body changed: %s -> %s", body1, body2)
-			}
-			if firstRID == "" || firstRID == secondRID {
-				t.Errorf("rid not regenerated: %q -> %q", firstRID, secondRID)
-			}
-			fake.ReplyOK(secondRID, map[string]any{
-				"id": fixedID, "send_at_ms": second["send_at_ms"], "state": "scheduled",
-			})
-			return
-		}
-		t.Error("timed out waiting for send retry")
-	}()
-
-	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
-	defer cancel()
-	res, err := c.Send(ctx, Target{Kind: "endpoint", ID: "b"}, Body{Enc: "utf8", Data: "hi"}, SendOptions{ID: fixedID, SendAt: &at})
-	<-done
-	if err != nil {
-		t.Fatalf("Send hung or failed: %v", err)
-	}
-	if res.ID != fixedID {
-		t.Fatalf("result id=%q want %q", res.ID, fixedID)
-	}
-	if firstRID == "" || secondRID == "" || firstRID == secondRID {
-		t.Fatalf("rids=%q/%q", firstRID, secondRID)
-	}
-}
-
 func TestResendKeepsIDAndSendAt(t *testing.T) {
 	fake := NewFakeTransport()
 	c := connectFake(t, fake)
@@ -3,7 +3,6 @@ package nixmsg
 import (
 	"context"
 	"encoding/json"
-	"errors"
 	"sync"
 	"sync/atomic"
 	"time"
@@ -30,8 +29,6 @@ type FakeTransport struct {
 	MaxFrameBytes     int
 	HelloDelay        time.Duration
 	ReceiveMaximumSet bool
-	// failSendLeft 接下来若干次 type=send 的 PublishUp 返回错误（仍记入 up）。
-	failSendLeft int
 }

 type fakeConnect struct {
@@ -68,13 +65,6 @@ func (f *FakeTransport) Start(ctx context.Context, cfg transportConfig) error {
 	return nil
 }

-// FailSendPublishN 让接下来 n 次 send 帧 PublishUp 失败（hello 等其它类型不受影响）。
-func (f *FakeTransport) FailSendPublishN(n int) {
-	f.mu.Lock()
-	f.failSendLeft = n
-	f.mu.Unlock()
-}
-
 func (f *FakeTransport) PublishUp(payload []byte) error {
 	f.mu.Lock()
 	cp := append([]byte(nil), payload...)
@@ -90,17 +80,6 @@ func (f *FakeTransport) PublishUp(payload []byte) error {
 	if auto && head.Type == "hello" && head.RID != "" {
 		f.replyHello(head.RID)
 	}
-	if head.Type == "send" {
-		f.mu.Lock()
-		fail := f.failSendLeft > 0
-		if fail {
-			f.failSendLeft--
-		}
-		f.mu.Unlock()
-		if fail {
-			return errors.New("publish failed")
-		}
-	}
 	return nil
 }

@@ -202,24 +202,19 @@ func (c *Client) dispatchSend(tr transport, item *sendItem, rid string, payload
 	if err := tr.PublishUp(payload); err != nil {
 		c.mu.Lock()
 		delete(c.pending, rid)
-		if item.epoch != epoch || !item.inflight {
-			c.mu.Unlock()
-			return
+		if item.epoch == epoch && item.inflight {
+			item.inflight = false
+			if c.inflight > 0 {
+				c.inflight--
+			}
+			if c.stopReconnect {
+				errStop := c.stopErrLocked()
+				c.mu.Unlock()
+				c.finishSend(item, SendResult{}, errStop)
+				return
+			}
 		}
-		item.inflight = false
-		if c.inflight > 0 {
-			c.inflight--
-		}
-		if c.stopReconnect {
-			errStop := c.stopErrLocked()
-			c.mu.Unlock()
-			c.finishSend(item, SendResult{}, errStop)
-			return
-		}
-		// 保留 id/body/send_at_ms，换新 rid 后继续泵，避免 Send 永久挂起。
-		c.regenerateSendLocked(item)
 		c.mu.Unlock()
-		c.drainSendQueue()
 		return
 	}