werkdock/internal/build/copy.go

package build

import (
	"fmt"
	"io"
	"os"
	"path/filepath"
	"syscall"

	"werkdock/internal/engine"
)

// CopyBackend builds in a full copy of the base rootfs, bound
// read-write into every RUN step (RFC 0006). It needs nothing from
// bwrap that 0.8 does not have: a writable root bind works in an
// unprivileged user namespace (measured), and no mount is made inside
// the sandbox, which could not make one anyway (CapEff is 0).
type CopyBackend struct{}

func (CopyBackend) Name() string { return "copy" }

// CachesSteps is false: without a union mount a cached step's result is
// a whole tree, so caching four steps of a 2 GiB image would keep four
// 2 GiB copies. The per-step cache waits for the overlay backend.
func (CopyBackend) CachesSteps() bool { return false }

// Open copies base to dir/rootfs.
func (CopyBackend) Open(base, dir string) (Tree, error) {
	rootfs := filepath.Join(dir, "rootfs")
	if err := copyTree(base, rootfs); err != nil {
		return nil, fmt.Errorf("copying the base image: %w", err)
	}
	return &copyTreeHandle{rootfs: rootfs}, nil
}

type copyTreeHandle struct {
	rootfs string
}

// Step hands out the same writable tree for every step: the copy
// backend has one tree, and a step's writes are simply in it.
func (t *copyTreeHandle) Step(int) (Step, error) {
	if t.rootfs == "" {
		return nil, fmt.Errorf("the work tree has been committed")
	}
	return copyStep{rootfs: t.rootfs}, nil
}

// Commit hands the tree over: it is already the flat rootfs the store
// wants, so the commit the caller then performs is a rename.
func (t *copyTreeHandle) Commit() (string, error) {
	if t.rootfs == "" {
		return "", fmt.Errorf("the work tree has already been committed")
	}
	rootfs := t.rootfs
	t.rootfs = ""
	return rootfs, nil
}

// Close has nothing to release: the tree lives in the caller's staging
// directory, which the caller removes when the build failed.
func (t *copyTreeHandle) Close() error { return nil }

type copyStep struct {
	rootfs string
}

func (s copyStep) Root(spec *engine.RunSpec) {
	spec.RootFS = s.rootfs
	spec.WritableRoot = true
}

func (s copyStep) Writable() string { return s.rootfs }

// copyTree duplicates a directory tree with real copies of the files.
//
// It is deliberately NOT store.copyTree, which hardlinks so that a
// second tag of one image costs no data — there both copies are
// read-only. Here a RUN step writes into the tree in place, and
// through a hardlink that write would land in the base image as well.
// A build needs a real copy; that copy is the cost of the copy backend
// (RFC 0006, Costs).
//
// Hardlinks *inside* the base are reproduced inside the copy (busybox
// links hundreds of names to one binary), which costs nothing and
// keeps the tree's own structure: those links point at the copy, never
// at the base.
func copyTree(src, dst string) error {
	linked := map[uint64]string{}
	return filepath.Walk(src, func(p string, fi os.FileInfo, err error) error {
		if err != nil {
			return err
		}
		rel, relErr := filepath.Rel(src, p)
		if relErr != nil {
			return relErr
		}
		target := filepath.Join(dst, rel)
		switch {
		case fi.IsDir():
			if err := os.MkdirAll(target, 0o700); err != nil {
				return err
			}
			// Owner-writable, whatever the image says: the steps have
			// to be able to add to every directory, and the finished
			// tree is bound read-only into every run anyway — as
			// store.applyLayer does for the same reason.
			return os.Chmod(target, fi.Mode().Perm()|0o700)
		case fi.Mode()&os.ModeSymlink != 0:
			link, linkErr := os.Readlink(p)
			if linkErr != nil {
				return linkErr
			}
			return os.Symlink(link, target)
		case fi.Mode().IsRegular():
			if ino, ok := hardlinkKey(fi); ok {
				if first, seen := linked[ino]; seen {
					return os.Link(first, target)
				}
				linked[ino] = target
			}
			return copyFile(p, target, fi.Mode().Perm())
		default:
			// Sockets, fifos, device nodes: not image content. /dev
			// comes from the host (RFC 0002) and store.applyLayer
			// skips them when unpacking, so a base image holds none.
			return nil
		}
	})
}

// hardlinkKey is the inode of a file that has more than one name, so
// that the second name becomes a link inside the copy.
func hardlinkKey(fi os.FileInfo) (uint64, bool) {
	st, ok := fi.Sys().(*syscall.Stat_t)
	if !ok || st.Nlink < 2 {
		return 0, false
	}
	return st.Ino, true
}

func copyFile(src, dst string, mode os.FileMode) error {
	in, err := os.Open(src)
	if err != nil {
		return err
	}
	defer in.Close()
	out, err := os.OpenFile(dst, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, mode)
	if err != nil {
		return err
	}
	if _, err := io.Copy(out, in); err != nil {
		out.Close()
		return err
	}
	if err := out.Close(); err != nil {
		return err
	}
	// O_CREATE applies the umask to mode, so set it explicitly, as the
	// layer unpacking does.
	return os.Chmod(dst, mode)
}