Replace overlay2 mount reexec with in-proc impl

Building off insights from the great work Cory Snider has been doing,
this replaces a reexec with a much lower overhead implementation which
performs the `Chddir` in a new goroutine that is locked to a specific
thread with CLONE_FS unshared.
The thread is thrown away afterwards and the Chdir does effectively the
same thing as what the reexec was being used for.

Signed-off-by: Brian Goff <cpuguy83@gmail.com>
This commit is contained in:
Brian Goff
2022-10-11 23:22:49 +00:00
parent 058010187b
commit 34f459423a

View File

@@ -4,94 +4,28 @@
package overlay2 // import "github.com/docker/docker/daemon/graphdriver/overlay2"
import (
"bytes"
"encoding/json"
"flag"
"fmt"
"os"
"runtime"
"github.com/docker/docker/pkg/reexec"
"golang.org/x/sys/unix"
)
func init() {
reexec.Register("docker-mountfrom", mountFromMain)
}
func fatal(err error) {
fmt.Fprint(os.Stderr, err)
os.Exit(1)
}
type mountOptions struct {
Device string
Target string
Type string
Label string
Flag uint32
}
func mountFrom(dir, device, target, mType string, flags uintptr, label string) error {
options := &mountOptions{
Device: device,
Target: target,
Type: mType,
Flag: uint32(flags),
Label: label,
}
chErr := make(chan error, 1)
cmd := reexec.Command("docker-mountfrom", dir)
w, err := cmd.StdinPipe()
if err != nil {
return fmt.Errorf("mountfrom error on pipe creation: %v", err)
}
go func() {
runtime.LockOSThread()
// Do not unlock this thread as the thread state cannot be restored
// We do not want go to re-use this thread for anything else.
output := bytes.NewBuffer(nil)
cmd.Stdout = output
cmd.Stderr = output
// reexec.Command() sets cmd.SysProcAttr.Pdeathsig on Linux, which
// causes the started process to be signaled when the creating OS thread
// dies. Ensure that the reexec is not prematurely signaled. See
// https://go.dev/issue/27505 for more information.
runtime.LockOSThread()
defer runtime.UnlockOSThread()
if err := cmd.Start(); err != nil {
w.Close()
return fmt.Errorf("mountfrom error on re-exec cmd: %v", err)
}
// write the options to the pipe for the untar exec to read
if err := json.NewEncoder(w).Encode(options); err != nil {
w.Close()
return fmt.Errorf("mountfrom json encode to pipe failed: %v", err)
}
w.Close()
if err := cmd.Wait(); err != nil {
return fmt.Errorf("mountfrom re-exec error: %v: output: %v", err, output)
}
return nil
}
// mountfromMain is the entry-point for docker-mountfrom on re-exec.
func mountFromMain() {
runtime.LockOSThread()
flag.Parse()
var options *mountOptions
if err := json.NewDecoder(os.Stdin).Decode(&options); err != nil {
fatal(err)
}
if err := os.Chdir(flag.Arg(0)); err != nil {
fatal(err)
}
if err := unix.Mount(options.Device, options.Target, options.Type, uintptr(options.Flag), options.Label); err != nil {
fatal(err)
}
os.Exit(0)
if err := unix.Unshare(unix.CLONE_FS); err != nil {
chErr <- err
return
}
if err := unix.Chdir(dir); err != nil {
chErr <- err
return
}
chErr <- unix.Mount(device, target, mType, flags, label)
}()
return <-chErr
}