container: implement autoroot as setup op

This code is useful beyond just pd behaviour, and implementing it this way also reduces IPC overhead.

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

2 files changed, 99 insertions, 0 deletions

diff --git a/container/autoroot.go b/container/autoroot.go
new file mode 100644
index 00000000..b6375b8c
--- /dev/null
+++ b/container/autoroot.go
@@ -0,0 +1,91 @@
+package container
+
+import (
+	"encoding/gob"
+	"fmt"
+	"os"
+	"path"
+	. "syscall"
+)
+
+func init() { gob.Register(new(AutoRootOp)) }
+
+// Root appends an [Op] that expands a directory into a toplevel bind mount mirror on container root.
+// This is not a generic setup op. It is implemented here to reduce ipc overhead.
+func (f *Ops) Root(host, prefix string, flags int) *Ops {
+	*f = append(*f, &AutoRootOp{host, prefix, flags, nil})
+	return f
+}
+
+type AutoRootOp struct {
+	Host, Prefix string
+	// passed through to bindMount
+	Flags int
+
+	// obtained during early;
+	// these wrap the underlying Op because BindMountOp is relatively complex,
+	// so duplicating that code would be unwise
+	resolved []Op
+}
+
+func (r *AutoRootOp) early(params *Params) error {
+	if !path.IsAbs(r.Host) {
+		return msg.WrapErr(EBADE, fmt.Sprintf("path %q is not absolute", r.Host))
+	}
+
+	if d, err := os.ReadDir(r.Host); err != nil {
+		return wrapErrSelf(err)
+	} else {
+		r.resolved = make([]Op, 0, len(d))
+		for _, ent := range d {
+			name := ent.Name()
+			if IsAutoRootBindable(name) {
+				op := &BindMountOp{
+					Source: path.Join(r.Host, name),
+					Target: "/" + name,
+					Flags:  r.Flags,
+				}
+				if err = op.early(params); err != nil {
+					return err
+				}
+				r.resolved = append(r.resolved, op)
+			}
+		}
+		return nil
+	}
+}
+
+func (r *AutoRootOp) apply(params *Params) error {
+	for _, op := range r.resolved {
+		msg.Verbosef("%s %s", op.prefix(), op)
+		if err := op.apply(params); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (r *AutoRootOp) Is(op Op) bool {
+	vr, ok := op.(*AutoRootOp)
+	return ok && ((r == nil && vr == nil) || (r != nil && vr != nil &&
+		r.Host == vr.Host && r.Prefix == vr.Prefix && r.Flags == vr.Flags))
+}
+func (*AutoRootOp) prefix() string { return "setting up" }
+func (r *AutoRootOp) String() string {
+	return fmt.Sprintf("auto root %q prefix %s flags %#x", r.Host, r.Prefix, r.Flags)
+}
+
+// IsAutoRootBindable returns whether a dir entry name is selected for AutoRoot.
+func IsAutoRootBindable(name string) bool {
+	switch name {
+	case "proc":
+	case "dev":
+	case "tmp":
+	case "mnt":
+	case "etc":
+
+	default:
+		return true
+	}
+	return false
+}
diff --git a/container/init.go b/container/init.go
index d4fce6e6..8b178d8c 100644
--- a/container/init.go
+++ b/container/init.go
@@ -121,6 +121,10 @@ func Init(prepare func(prefix string), setVerbose func(verbose bool)) {
 		log.Fatalf("cannot make / rslave: %v", err)
 	}
 
+	/* early is called right before pivot_root into intermediate root;
+	this step is mostly for gathering information that would otherwise be difficult to obtain
+	via library functions after pivot_root, and implementations are expected to avoid changing
+	the state of the mount namespace */
 	for i, op := range *params.Ops {
 		if op == nil {
 			log.Fatalf("invalid op %d", i)
@@ -159,6 +163,10 @@ func Init(prepare func(prefix string), setVerbose func(verbose bool)) {
 		log.Fatalf("%v", err)
 	}
 
+	/* apply is called right after pivot_root and entering the new root;
+	this step sets up the container filesystem, and implementations are expected to keep the host root
+	and sysroot mount points intact but otherwise can do whatever they need to;
+	chdir is allowed but discouraged */
 	for i, op := range *params.Ops {
 		// ops already checked during early setup
 		msg.Verbosef("%s %s", op.prefix(), op)