Add retry logic during aufs unmount.

Treat EBUSY as a transient error and retry. Also stop ignoring unmount errors.

Signed-off-by: Anusha Ragunathan <anusha@docker.com>
(cherry picked from commit 0e539fec33)
Signed-off-by: Victor Vieux <vieux@docker.com>

daemon/graphdriver/aufs/aufs.go

@@ -33,6 +33,7 @@ import (
 	"strings"
 	"sync"
 	"syscall"
+	"time"
 
 	"github.com/Sirupsen/logrus"
 	"github.com/vbatts/tar-split/tar/storage"
@@ -264,6 +265,39 @@ func (a *Driver) createDirsFor(id string) error {
 	return nil
 }
 
+// Helper function to debug EBUSY errors on remove.
+func debugEBusy(mountPath string) (out []string, err error) {
+	// lsof is not part of GNU coreutils. This is a best effort
+	// attempt to detect offending processes.
+	c := exec.Command("lsof")
+
+	r, err := c.StdoutPipe()
+	if err != nil {
+		return nil, fmt.Errorf("Assigning pipes failed with %v", err)
+	}
+
+	if err := c.Start(); err != nil {
+		return nil, fmt.Errorf("Starting %s failed with %v", c.Path, err)
+	}
+
+	defer func() {
+		waiterr := c.Wait()
+		if waiterr != nil && err == nil {
+			err = fmt.Errorf("Waiting for %s failed with %v", c.Path, waiterr)
+		}
+	}()
+
+	sc := bufio.NewScanner(r)
+	for sc.Scan() {
+		entry := sc.Text()
+		if strings.Contains(entry, mountPath) {
+			out = append(out, entry, "\n")
+		}
+	}
+
+	return out, nil
+}
+
 // Remove will unmount and remove the given id.
 func (a *Driver) Remove(id string) error {
 	a.pathCacheLock.Lock()
@@ -272,9 +306,35 @@ func (a *Driver) Remove(id string) error {
 	if !exists {
 		mountpoint = a.getMountpoint(id)
 	}
-	if err := a.unmount(mountpoint); err != nil {
-		// no need to return here, we can still try to remove since the `Rename` will fail below if still mounted
-		logrus.Debugf("aufs: error while unmounting %s: %v", mountpoint, err)
+
+	var retries int
+	for {
+		mounted, err := a.mounted(mountpoint)
+		if err != nil {
+			return err
+		}
+		if !mounted {
+			break
+		}
+
+		if err := a.unmount(mountpoint); err != nil {
+			if err != syscall.EBUSY {
+				return fmt.Errorf("aufs: unmount error: %s: %v", mountpoint, err)
+			}
+			if retries >= 5 {
+				out, debugErr := debugEBusy(mountpoint)
+				if debugErr == nil {
+					logrus.Warnf("debugEBusy returned %v", out)
+				}
+				return fmt.Errorf("aufs: unmount error after retries: %s: %v", mountpoint, err)
+			}
+			// If unmount returns EBUSY, it could be a transient error. Sleep and retry.
+			retries++
+			logrus.Warnf("unmount failed due to EBUSY: retry count: %d", retries)
+			time.Sleep(100 * time.Millisecond)
+			continue
+		}
+		break
 	}
 
 	// Atomically remove each directory in turn by first moving it out of the
@@ -282,6 +342,13 @@ func (a *Driver) Remove(id string) error {
 	// the whole tree.
 	tmpMntPath := path.Join(a.mntPath(), fmt.Sprintf("%s-removing", id))
 	if err := os.Rename(mountpoint, tmpMntPath); err != nil && !os.IsNotExist(err) {
+		if err == syscall.EBUSY {
+			logrus.Warnf("os.Rename err due to EBUSY")
+			out, debugErr := debugEBusy(mountpoint)
+			if debugErr == nil {
+				logrus.Warnf("debugEBusy returned %v", out)
+			}
+		}
 		return err
 	}
 	defer os.RemoveAll(tmpMntPath)