container: set CLOEXEC via close_range

This is guarded behind the close_range build tag for now.

Signed-off-by: Ophestra <cat@gensokyo.uk>

3 files changed, 52 insertions, 20 deletions

diff --git a/container/container.go b/container/container.go
index 7d8837a3..86bb3b36 100644
--- a/container/container.go
+++ b/container/container.go
@@ -186,31 +186,24 @@ var (
 	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.
+// 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)
-		}
-	})
+	closeOnExecOnce.Do(func() { closeOnExecErr = doCloseOnExec() })
 
 	if closeOnExecErr == nil {
 		return nil
 	}
-	return &StartError{Fatal: true, Step: "set FD_CLOEXEC on all open files", Err: closeOnExecErr, Passthrough: true}
+	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.
diff --git a/container/syscall_close_range.go b/container/syscall_close_range.go
new file mode 100644
index 00000000..c4d6a051
--- /dev/null
+++ b/container/syscall_close_range.go
@@ -0,0 +1,11 @@
+//go:build close_range
+
+package container
+
+import "hakurei.app/ext"
+
+// doCloseOnExec implements ensureCloseOnExec by calling CloseRange with
+// CLOSE_RANGE_CLOEXEC.
+func doCloseOnExec() error {
+	return ext.CloseRange(0, ext.MaxUint, ext.CLOSE_RANGE_CLOEXEC)
+}
diff --git a/container/syscall_range_proc.go b/container/syscall_range_proc.go
new file mode 100644
index 00000000..62b369f0
--- /dev/null
+++ b/container/syscall_range_proc.go
@@ -0,0 +1,28 @@
+//go:build !close_range
+
+package container
+
+import (
+	"os"
+	"strconv"
+	"syscall"
+
+	"hakurei.app/container/fhs"
+)
+
+// doCloseOnExec implements ensureCloseOnExec by ranging over proc_pid_fd(5).
+func doCloseOnExec() error {
+	entries, err := os.ReadDir(fhs.ProcSelf + "fd/")
+	if err != nil {
+		return err
+	}
+
+	var fd int
+	for _, ent := range entries {
+		if fd, err = strconv.Atoi(ent.Name()); err != nil {
+			return err // not reached
+		}
+		syscall.CloseOnExec(fd)
+	}
+	return nil
+}