Fix race between client's close and write
After commit 506f3a7 which removed clientConn mutex around
conn.sendPacket, it is now possible for conn.Close and conn.Write to
race (https://travis-ci.org/pkg/sftp/jobs/219048782#L1838). The problem
is really hard to reproduce, but I believe this patch fixes it.
Now clientConn.Mutex protects 'inflight' map only, and conn.Mutex
serializes connection writes/close.
Signed-off-by: Pavel Borzenkov <pavel.borzenkov@gmail.com>
diff --git a/conn.go b/conn.go
index 4b0d834..d9e1095 100644
--- a/conn.go
+++ b/conn.go
@@ -56,9 +56,9 @@
// appropriate channel.
func (c *clientConn) recv() error {
defer func() {
- c.Lock()
+ c.conn.Lock()
c.conn.Close()
- c.Unlock()
+ c.conn.Unlock()
}()
for {
typ, data, err := c.recvPacket()
