werkdock/internal/api/archive.go

package api

import (
	"archive/tar"
	"compress/gzip"
	"fmt"
	"io"
	"net/http"
	"os"
	"path"
	"path/filepath"
	"strings"

	"werkdock/internal/engine"
)

// maxArchiveEntries bounds what one PUT may unpack. Every top-level
// entry becomes a bind mount and therefore two words of bwrap's argv;
// a client that wants to populate a whole tree wants an image, not a
// copy.
const maxArchiveEntries = 64

// containerArchivePut is docker's PUT /containers/{id}/archive, the
// call behind withCopyFileToContainer: a tar stream unpacked at the
// directory named by ?path=.
//
// Werkdock does it differently underneath, because it has to. The
// rootfs is the image's own directory in the store, shared by every
// instance and bound read-only; unpacking into it would change the
// image for everyone. So the tar is unpacked into the instance
// directory and bound into the sandbox — in one of two ways, and the
// image decides which:
//
//   - The destination is empty in the image, or absent: the unpacked
//     directory is bound over it whole. Nothing is hidden, because
//     there was nothing there, and the image gets one empty directory
//     as a mountpoint.
//   - The image already fills the destination (/etc, say): each
//     unpacked file is bound to its own path instead, so the image's
//     files stay visible beside the copied ones — docker's merge. The
//     price is a mountpoint per file in the shared image directory,
//     empty files that other instances of the image will see. Werkdock
//     pre-creates mountpoints there for every bind (EnsureMountpoints);
//     this is that same wart, not a new one.
//
// A copied *directory* aimed at one the image fills cannot be
// expressed either way — binding it would hide what is inside — and is
// refused with the reason rather than done quietly.
//
// path=/ is the ordinary case, not a special one: docker's clients
// send the whole path in the tar entries and / as the destination. The
// rootfs is never empty, so that request always takes the second way
// and lands one level down, where it belongs.
//
// The mounts are fixed when the sandbox starts, so this works on a
// created instance and is refused on a running one, with the reason.
func (s *Server) containerArchivePut(w http.ResponseWriter, r *http.Request) {
	inst, ok := s.instanceOf(w, r)
	if !ok {
		return
	}
	if inst.Running() {
		writeError(w, http.StatusNotImplemented,
			"cannot copy into a running instance: a sandbox's mounts are fixed when it starts, "+
				"and werkdock has no writable container layer to put the file in. "+
				"Copy before start (withCopyFileToContainer does), or write through a volume")
		return
	}
	dest := r.URL.Query().Get("path")
	if dest == "" {
		writeError(w, http.StatusBadRequest, "path is required")
		return
	}
	if !strings.HasPrefix(dest, "/") {
		dest = "/" + dest
	}
	dest = path.Clean(dest)

	body, err := maybeGunzip(r.Body)
	if err != nil {
		writeError(w, http.StatusBadRequest, "%v", err)
		return
	}
	root := filepath.Join(inst.Dir, "archive", filepath.FromSlash(strings.TrimPrefix(dest, "/")))
	if err := os.MkdirAll(root, 0o755); err != nil {
		writeError(w, http.StatusInternalServerError, "%v", err)
		return
	}
	names, err := unpack(tar.NewReader(body), root)
	if err != nil {
		writeError(w, http.StatusBadRequest, "%v", err)
		return
	}

	spec := inst.Config.Spec
	if imageFills(spec.RootFS, dest) {
		// The image has files here; bind the copies one by one so its
		// own stay visible.
		for _, name := range names {
			source := filepath.Join(root, name)
			target := path.Join(dest, name)
			info, err := os.Stat(source)
			if err != nil {
				writeError(w, http.StatusInternalServerError, "%v", err)
				return
			}
			if info.IsDir() && imageFills(spec.RootFS, target) {
				writeError(w, http.StatusNotImplemented,
					"cannot copy a directory onto %s: the image already has files there, and werkdock "+
						"would have to hide them — it binds the copy in rather than writing into the shared "+
						"read-only image. Copy the files themselves, or use a volume", target)
				return
			}
			spec.Mounts = addMount(spec.Mounts, engine.Mount{Mode: engine.MountBind, Source: source, Dest: target})
		}
	} else {
		// Nothing of the image's is at stake; one mount carries the lot,
		// and a second copy to the same path lands in the same directory.
		spec.Mounts = addMount(spec.Mounts, engine.Mount{Mode: engine.MountBind, Source: root, Dest: dest})
	}
	if err := engine.EnsureMountpoints(spec); err != nil {
		writeError(w, http.StatusInternalServerError, "%v", err)
		return
	}
	inst.Config.Spec = spec
	if err := inst.SaveConfig(); err != nil {
		writeError(w, http.StatusInternalServerError, "%v", err)
		return
	}
	s.logf("copied %d entries into %s at %s", len(names), inst.ShortID(), dest)
	w.WriteHeader(http.StatusOK)
}

// unpack writes the tar below root and returns the top-level names, in
// the order they arrived. Only regular files and directories: a
// symlink or a device node in a copied archive would mean something
// werkdock cannot honour, and saying so beats unpacking half of it.
func unpack(tr *tar.Reader, root string) ([]string, error) {
	var names []string
	seen := map[string]bool{}
	count := 0
	for {
		header, err := tr.Next()
		if err == io.EOF {
			break
		}
		if err != nil {
			return nil, fmt.Errorf("reading the archive: %w", err)
		}
		name := path.Clean(strings.TrimPrefix(header.Name, "./"))
		if name == "." || name == "/" {
			continue
		}
		target, err := safeJoin(root, name)
		if err != nil {
			return nil, err
		}
		count++
		if count > maxArchiveEntries {
			return nil, fmt.Errorf("the archive holds more than %d entries; werkdock binds each one "+
				"into the sandbox and will not build an argv that long", maxArchiveEntries)
		}
		switch header.Typeflag {
		case tar.TypeDir:
			if err := os.MkdirAll(target, 0o755); err != nil {
				return nil, err
			}
		case tar.TypeReg:
			if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
				return nil, err
			}
			f, err := os.OpenFile(target, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, os.FileMode(header.Mode).Perm())
			if err != nil {
				return nil, err
			}
			if _, err := io.Copy(f, tr); err != nil {
				f.Close()
				return nil, err
			}
			if err := f.Close(); err != nil {
				return nil, err
			}
		default:
			return nil, fmt.Errorf("%s is a %s; werkdock copies regular files and directories only, "+
				"because everything else would have to mean something inside the sandbox that it cannot promise",
				header.Name, typeName(header.Typeflag))
		}
		top := name
		if i := strings.Index(top, "/"); i >= 0 {
			top = top[:i]
		}
		if !seen[top] {
			seen[top] = true
			names = append(names, top)
		}
	}
	if len(names) == 0 {
		return nil, fmt.Errorf("the archive is empty")
	}
	return names, nil
}

func typeName(flag byte) string {
	switch flag {
	case tar.TypeSymlink:
		return "symlink"
	case tar.TypeLink:
		return "hard link"
	case tar.TypeChar, tar.TypeBlock:
		return "device node"
	case tar.TypeFifo:
		return "fifo"
	}
	return "special entry"
}

// imageFills reports whether the image has anything at dest, which is
// what decides between the two ways of binding a copy in.
func imageFills(rootfs, dest string) bool {
	target := filepath.Join(rootfs, filepath.FromSlash(strings.TrimPrefix(dest, "/")))
	entries, err := os.ReadDir(target)
	return err == nil && len(entries) > 0
}

// addMount keeps one mount per destination, so copying twice to the
// same place does not bind it twice.
func addMount(mounts []engine.Mount, m engine.Mount) []engine.Mount {
	for _, existing := range mounts {
		if existing.Dest == m.Dest {
			return mounts
		}
	}
	return append(mounts, m)
}

// safeJoin refuses an archive that tries to write outside its root.
func safeJoin(root, name string) (string, error) {
	target := filepath.Join(root, filepath.FromSlash(name))
	if target != root && !strings.HasPrefix(target, root+string(os.PathSeparator)) {
		return "", fmt.Errorf("the archive entry %q leads outside the destination", name)
	}
	return target, nil
}

// maybeGunzip accepts the compressed tar docker's clients may send.
func maybeGunzip(r io.Reader) (io.Reader, error) {
	buffered := make([]byte, 2)
	n, err := io.ReadFull(r, buffered)
	if err == io.EOF || err == io.ErrUnexpectedEOF {
		return strings.NewReader(""), nil
	}
	if err != nil {
		return nil, err
	}
	joined := io.MultiReader(strings.NewReader(string(buffered[:n])), r)
	if buffered[0] == 0x1f && buffered[1] == 0x8b {
		return gzip.NewReader(joined)
	}
	return joined, nil
}