container: set FD_CLOEXEC on all open files
All checks were successful
Test / Create distribution (push) Successful in 29s
Test / Sandbox (race detector) (push) Successful in 40s
Test / Hakurei (race detector) (push) Successful in 46s
Test / Hakurei (push) Successful in 47s
Test / Sandbox (push) Successful in 44s
Test / Hpkg (push) Successful in 43s
Test / Flake checks (push) Successful in 1m31s

While fd created from this side always has the FD_CLOEXEC flag, the same is not true for files left open by the parent. This change prevents those files from leaking into the container.

Signed-off-by: Ophestra <cat@gensokyo.uk>
This commit is contained in:
2025-11-12 00:18:11 +09:00
parent 9dec9dbc4b
commit ac34635890
3 changed files with 65 additions and 21 deletions

View File

@@ -11,6 +11,7 @@ import (
"os/exec"
"runtime"
"strconv"
"sync"
. "syscall"
"time"
@@ -143,11 +144,18 @@ func (e *StartError) Error() string {
// Message returns a user-facing error message.
func (e *StartError) Message() string {
if e.Passthrough {
var (
numError *strconv.NumError
)
switch {
case errors.As(e.Err, new(*os.PathError)),
errors.As(e.Err, new(*os.SyscallError)):
return "cannot " + e.Err.Error()
case errors.As(e.Err, &numError) && numError != nil:
return "cannot parse " + strconv.Quote(numError.Num) + ": " + numError.Err.Error()
default:
return e.Err.Error()
}
@@ -158,6 +166,39 @@ func (e *StartError) Message() string {
return "cannot " + e.Error()
}
// for ensureCloseOnExec
var (
closeOnExecOnce sync.Once
closeOnExecErr error
)
// ensureCloseOnExec ensures all currently open file descriptors have the syscall.FD_CLOEXEC flag set.
// This is only ran once as it is intended to handle files left open by the parent, and any file opened
// on this side should already have syscall.FD_CLOEXEC set.
func ensureCloseOnExec() error {
closeOnExecOnce.Do(func() {
const fdPrefixPath = "/proc/self/fd/"
var entries []os.DirEntry
if entries, closeOnExecErr = os.ReadDir(fdPrefixPath); closeOnExecErr != nil {
return
}
var fd int
for _, ent := range entries {
if fd, closeOnExecErr = strconv.Atoi(ent.Name()); closeOnExecErr != nil {
break // not reached
}
CloseOnExec(fd)
}
})
if closeOnExecErr == nil {
return nil
}
return &StartError{Fatal: true, Step: "set FD_CLOEXEC on all open files", Err: closeOnExecErr, Passthrough: true}
}
// Start starts the container init. The init process blocks until Serve is called.
func (p *Container) Start() error {
if p == nil || p.cmd == nil ||
@@ -168,6 +209,10 @@ func (p *Container) Start() error {
return errors.New("container: already started")
}
if err := ensureCloseOnExec(); err != nil {
return err
}
// map to overflow id to work around ownership checks
if p.Uid < 1 {
p.Uid = OverflowUid(p.msg)