fix(v2): 超退守卫事务化 CreateRefundGuarded(锁单行+求和+插入单事务)+ sqlite DSN 补 _txlock=immediate

Task 3 计划的 RefundSum→比较→CreateRefund 两段式裸读写在并发下会超退(两个请求都读到
reserved=0、都通过、都插入)。仿 store/order.go::MarkAttemptPaid 的模式,把锁订单行
(clause.Locking FOR UPDATE,MySQL 真锁/SQLite 由 glebarez 静默丢弃)+ 求和 + 校验 + 插入
收进单个 s.db.Transaction。SQLite 侧真正的串行化靠 DSN _txlock=immediate(BEGIN IMMEDIATE
在事务开始就抢写锁)+ busy_timeout,补进 testdb.go 与 main.go 的 sqlite DSN 构造。
This commit is contained in:
wangjia
2026-07-10 17:08:25 +08:00
parent 338e28d386
commit aa14204640
4 changed files with 201 additions and 2 deletions
+6 -1
View File
@@ -20,7 +20,12 @@ var testDBCounter int64
func OpenTestDB(t *testing.T) *gorm.DB {
t.Helper()
n := atomic.AddInt64(&testDBCounter, 1)
dsn := fmt.Sprintf("file:testdb_%d?mode=memory&cache=shared", n)
// _txlock=immediate: BEGIN IMMEDIATE 让每个事务一开始就抢库级写锁(而非默认
// BEGIN DEFERRED 推迟到首条写语句才抢),使并发事务在读阶段就串行化——守卫类
// 事务(锁行/求和/校验/插入,见 RefundStore.CreateRefundGuarded)依赖此语义,
// 呼应 pangolin 的 SQLite DSN 约定(server/internal/db/db.go)。
// _pragma=busy_timeout(5000): 抢不到写锁时等待重试而非立即 SQLITE_BUSY 报错。
dsn := fmt.Sprintf("file:testdb_%d?mode=memory&cache=shared&_txlock=immediate&_pragma=busy_timeout(5000)", n)
db, err := gorm.Open(sqlite.Open(dsn),
&gorm.Config{Logger: logger.Default.LogMode(logger.Silent), TranslateError: true})
if err != nil {
+58
View File
@@ -6,6 +6,7 @@ import (
"time"
"gorm.io/gorm"
"gorm.io/gorm/clause"
"github.com/wangjia/pay/internal/model"
)
@@ -16,6 +17,11 @@ type RefundStore struct{ db *gorm.DB }
func NewRefundStore(db *gorm.DB) *RefundStore { return &RefundStore{db: db} }
// CreateRefund is a plain insert with no over-refund guard — safe only for
// callers that already hold their own serialization (e.g. tests, backfills).
// Money-path callers (退款下单) MUST use CreateRefundGuarded instead: a bare
// RefundSum(...)-then-CreateRefund round trip races under concurrent refund
// requests (both read reserved=0, both pass, both insert → over-refund).
func (s *RefundStore) CreateRefund(r *model.Refund) error {
if r.RefundID == "" {
return fmt.Errorf("store.CreateRefund: empty RefundID")
@@ -26,6 +32,58 @@ func (s *RefundStore) CreateRefund(r *model.Refund) error {
return nil
}
// CreateRefundGuarded 在单事务内完成 锁定订单行 → 求和(非failed)→ 超退校验 → 插入。
// 返回 (false, nil) 表示超退拒绝;并发同单退款经订单行锁串行化。
//
// 订单行锁 clause.Locking{Strength: "UPDATE"} 在 MySQL 上是真正的 SELECT ... FOR
// UPDATE 行锁;SQLite 不支持行级锁,glebarez/sqlite 的 "FOR" ClauseBuilder 会把它
// 静默丢弃(不报错、也不生效)。SQLite 侧真正的串行化来自 DSN 的
// _txlock=immediate(见 internal/model/testdb.go / main.go initDB):BEGIN IMMEDIATE
// 让事务一开始就抢库级写锁,第二个并发事务的 BEGIN 会阻塞到第一个提交/回滚为止,
// 效果等价于「锁行」。没有 _txlock=immediate 时 SQLite 默认 BEGIN DEFERRED,写锁
// 推迟到第一次写语句才抢,两个事务的读阶段(求和)仍可能交错,防不住超退。
func (s *RefundStore) CreateRefundGuarded(r *model.Refund, paidAmountMinor int64) (bool, error) {
if r.RefundID == "" {
return false, fmt.Errorf("store.CreateRefundGuarded: empty RefundID")
}
if r.OutTradeNo == "" {
return false, fmt.Errorf("store.CreateRefundGuarded: empty OutTradeNo")
}
var ok bool
err := s.db.Transaction(func(tx *gorm.DB) error {
var o model.OrderV2
if err := tx.Clauses(clause.Locking{Strength: "UPDATE"}).
Where("out_trade_no = ?", r.OutTradeNo).First(&o).Error; err != nil {
if errors.Is(err, gorm.ErrRecordNotFound) {
return ErrOrderNotFound
}
return fmt.Errorf("lock order: %w", err)
}
var reserved int64
q := tx.Model(&model.Refund{}).
Where("out_trade_no = ? AND status <> ?", r.OutTradeNo, model.RefundFailed)
if err := q.Select("COALESCE(SUM(amount_minor),0)").Scan(&reserved).Error; err != nil {
return fmt.Errorf("sum refunds: %w", err)
}
if reserved+r.AmountMinor > paidAmountMinor {
ok = false
return nil
}
if err := tx.Create(r).Error; err != nil {
return err
}
ok = true
return nil
})
if err != nil {
return false, fmt.Errorf("store.CreateRefundGuarded: %w", err)
}
return ok, nil
}
func (s *RefundStore) GetRefund(refundID string) (*model.Refund, error) {
var r model.Refund
if err := s.db.Where("refund_id = ?", refundID).First(&r).Error; err != nil {
+122
View File
@@ -1,9 +1,16 @@
package store
import (
"errors"
"fmt"
"sync"
"testing"
"time"
"github.com/glebarez/sqlite"
"gorm.io/gorm"
"gorm.io/gorm/logger"
"github.com/wangjia/pay/internal/model"
)
@@ -103,3 +110,118 @@ func TestListManualPending(t *testing.T) {
t.Fatalf("manual pending = %+v, %v", list, err)
}
}
// TestCreateRefundGuardedSequential 顺序场景:60 通过、再 50 超退拒(60+50=110>100)、
// 再 40 通过(60+40=100 打满),验证守卫在非并发路径下的基本判定与求和口径(非 failed)。
func TestCreateRefundGuardedSequential(t *testing.T) {
db := model.OpenTestDB(t)
os := NewOrderStore(db)
rs := NewRefundStore(db)
seedPaid(t, os, "PAY-G1", 10000) // 100.00
ok, err := rs.CreateRefundGuarded(&model.Refund{
RefundID: "g1-a", OutTradeNo: "PAY-G1", AttemptProviderRef: "PAY-G1",
AmountMinor: 6000, Currency: "CNY", Status: model.RefundProcessing, InitiatedBy: "business",
}, 10000)
if err != nil || !ok {
t.Fatalf("g1-a: ok=%v err=%v want true,nil", ok, err)
}
ok, err = rs.CreateRefundGuarded(&model.Refund{
RefundID: "g1-b", OutTradeNo: "PAY-G1", AttemptProviderRef: "PAY-G1",
AmountMinor: 5000, Currency: "CNY", Status: model.RefundProcessing, InitiatedBy: "business",
}, 10000)
if err != nil || ok {
t.Fatalf("g1-b: ok=%v err=%v want false,nil(超退应拒)", ok, err)
}
ok, err = rs.CreateRefundGuarded(&model.Refund{
RefundID: "g1-c", OutTradeNo: "PAY-G1", AttemptProviderRef: "PAY-G1",
AmountMinor: 4000, Currency: "CNY", Status: model.RefundProcessing, InitiatedBy: "business",
}, 10000)
if err != nil || !ok {
t.Fatalf("g1-c: ok=%v err=%v want true,nil", ok, err)
}
sum, err := rs.RefundSum("PAY-G1", model.RefundRequested, model.RefundProcessing, model.RefundManualPending, model.RefundSucceeded)
if err != nil || sum != 10000 {
t.Fatalf("reserved sum = %d, err=%v want 10000", sum, err)
}
// 被拒的 g1-b 不应落库
if _, err := rs.GetRefund("g1-b"); !errors.Is(err, ErrRefundNotFound) {
t.Fatalf("g1-b 应未落库,GetRefund err=%v", err)
}
}
// openFileGuardedDB 开一个 t.TempDir 下的文件型 sqlite(而非 model.OpenTestDB 的
// in-memory cache=shared),确保并发 goroutine 走真正独立的连接/文件锁路径来
// 验证 CreateRefundGuarded 的串行化 —— DSN 同样带 _txlock=immediate。
func openFileGuardedDB(t *testing.T) *gorm.DB {
t.Helper()
dsn := fmt.Sprintf("file:%s/guard.db?_txlock=immediate&_pragma=busy_timeout(5000)", t.TempDir())
db, err := gorm.Open(sqlite.Open(dsn),
&gorm.Config{Logger: logger.Default.LogMode(logger.Silent), TranslateError: true})
if err != nil {
t.Fatalf("open file guard db: %v", err)
}
if err := db.AutoMigrate(&model.OrderV2{}, &model.Attempt{}, &model.Account{}, &model.Refund{},
&model.WebhookDelivery{}, &model.Product{}, &model.ProductPrice{}); err != nil {
t.Fatalf("migrate: %v", err)
}
sqlDB, _ := db.DB()
t.Cleanup(func() { _ = sqlDB.Close() })
return db
}
// TestCreateRefundGuardedConcurrentExactlyOneWins 两个 goroutine 同时对同一笔
// 100.00 已付订单各申请 60.00 退款(合计 120 > 100,必超退)。事务化守卫(锁订单行 +
// 事务内求和 + 校验 + 插入)必须让恰好一个成功、另一个被拒 —— 若把守卫换回「先
// RefundSum 再 CreateRefund」的裸读写两段式(Task 3 计划里的朴素调用法),两个
// goroutine 都可能读到 reserved=0、都通过校验、都插入成功,变成两个 true(超退)。
// 本测试即用来钉住这个回归:TDD 上手动验证过 —— 把 CreateRefundGuarded 换成裸
// RefundSum(...)+CreateRefund(...) 两步、不用事务/锁,本测试会因为「两个 true」而
// FAIL(count==2),证明测试确实在防这一类竞态,而不是摆设。
func TestCreateRefundGuardedConcurrentExactlyOneWins(t *testing.T) {
db := openFileGuardedDB(t)
os := NewOrderStore(db)
rs := NewRefundStore(db)
seedPaid(t, os, "PAY-G2", 10000) // 100.00
var wg sync.WaitGroup
results := make([]bool, 2)
errs := make([]error, 2)
ids := []string{"g2-a", "g2-b"}
for i := 0; i < 2; i++ {
wg.Add(1)
go func(i int) {
defer wg.Done()
ok, err := rs.CreateRefundGuarded(&model.Refund{
RefundID: ids[i], OutTradeNo: "PAY-G2", AttemptProviderRef: "PAY-G2",
AmountMinor: 6000, Currency: "CNY", Status: model.RefundProcessing, InitiatedBy: "business",
}, 10000)
results[i] = ok
errs[i] = err
}(i)
}
wg.Wait()
trueCount := 0
for i, ok := range results {
// 允许 SQLite 忙锁重试耗尽等基础设施性错误(busy_timeout 已设为 5s 通常不会
// 触发),但不允许"两个都成功"的超退;若出现非预期错误直接报失败方便定位。
if errs[i] != nil {
t.Fatalf("goroutine %d unexpected error: %v", i, errs[i])
}
if ok {
trueCount++
}
}
if trueCount != 1 {
t.Fatalf("trueCount = %d, want 1 (results=%v)", trueCount, results)
}
sum, err := rs.RefundSum("PAY-G2", model.RefundRequested, model.RefundProcessing, model.RefundManualPending, model.RefundSucceeded)
if err != nil || sum != 6000 {
t.Fatalf("reserved sum = %d, err=%v want 6000(仅一笔落库)", sum, err)
}
}
+15 -1
View File
@@ -3,6 +3,7 @@ package main
import (
"errors"
"log"
"strings"
"time"
"github.com/gin-gonic/gin"
@@ -84,7 +85,7 @@ func initDB() *gorm.DB {
switch config.C.Database.Driver {
case "sqlite", "":
db, err := gorm.Open(sqlite.Open(config.C.Database.DSN), gormCfg)
db, err := gorm.Open(sqlite.Open(sqliteDSN(config.C.Database.DSN)), gormCfg)
if err != nil {
log.Fatalf("连接 sqlite 失败: %v", err)
}
@@ -98,6 +99,19 @@ func initDB() *gorm.DB {
}
}
// sqliteDSN 补 _txlock=immediate(BEGIN IMMEDIATE 立即抢库级写锁,串行化并发事务的
// 读阶段——RefundStore.CreateRefundGuarded 等守卫类事务依赖此语义)+
// busy_timeout(5000)(抢不到锁时等待重试而非立即报 SQLITE_BUSY)。呼应 pangolin
// 的 SQLite DSN 约定(server/internal/db/db.go)。config.yaml 里的 dsn 是纯文件路径
// (如 "pay.db"),这里补上查询串;若用户已经自带 "?",追加而不是覆盖。
func sqliteDSN(raw string) string {
extra := "_txlock=immediate&_pragma=busy_timeout(5000)"
if strings.Contains(raw, "?") {
return raw + "&" + extra
}
return raw + "?" + extra
}
func autoMigrate(db *gorm.DB) {
if err := db.AutoMigrate(
&model.Merchant{},