fix: 修复raceDial竞态条件导致daemon死锁无法连接

raceDial在首个连接成功后cancel剩余拨号goroutine时,dial goroutine
的select有两个同时就绪的case(缓冲channel可发送 vs raceCtx.Done()),
Go随机选择可能导致结果未发送,使drain循环永久阻塞,进而导致daemon
持有d.mu死锁,所有IPC命令(stats/stop/新start)全部阻塞。

DNS解析返回IPv4+IPv6双地址时触发此bug,约50%概率死锁。

修复:移除select,直接发送到缓冲channel(容量=len(ips)永不阻塞),
多余连接由drain循环负责关闭。
This commit is contained in:
2026-07-09 23:19:46 +08:00
parent e61d4fedd8
commit 469106d502
2 changed files with 142 additions and 8 deletions
+9 -8
View File
@@ -99,13 +99,13 @@ func raceDial(ctx context.Context, network string, ips []string, port string) (n
go func(target string) {
d := &net.Dialer{}
c, err := d.DialContext(raceCtx, network, net.JoinHostPort(target, port))
select {
case resultCh <- result{conn: c, err: err}:
case <-raceCtx.Done():
if c != nil {
c.Close()
}
}
// Always send the result: resultCh is buffered (len(ips)),
// so this never blocks. Do NOT use a select with
// <-raceCtx.Done() here - when the context is cancelled
// both cases would be ready and Go's random select pick
// could skip the send, causing the drain loop below to
// block forever.
resultCh <- result{conn: c, err: err}
}(ip)
}
@@ -117,7 +117,8 @@ func raceDial(ctx context.Context, network string, ips []string, port string) (n
// any late successful connections.
cancel()
for j := i + 1; j < len(ips); j++ {
if late := <-resultCh; late.conn != nil {
late := <-resultCh
if late.conn != nil {
late.conn.Close()
}
}