path: root/internal/apply/fs.go

// SPDX-License-Identifier: GPL-3.0-or-later

package apply

import (
	"errors"
	"fmt"
	"io"
	"io/fs"
	"os"
	"path/filepath"
	"strings"
	"syscall"
)

// copyFile copies src to dst by streaming its content through a temporary
// file created in dst's directory, syncing it, then renaming it into place.
// dst must not already exist. Mode and modification time are preserved from
// src. On any failure the temporary file is removed and neither src nor a
// pre-existing dst is touched.
//
// This is not internal/config's replaceFile reused: that helper resolves a
// destination symlink and overwrites a file that is already there, and it
// takes the whole replacement as a []byte. copy's destination must not exist
// beforehand, and a []byte cannot stand in for a file that may be many
// gigabytes, so this is a separate, streaming equivalent kept local to
// internal/apply rather than shared with internal/config.
func copyFile(src, dst string) error {
	in, err := os.Open(src)
	if err != nil {
		return err
	}
	defer in.Close()

	fi, err := in.Stat()
	if err != nil {
		return err
	}

	dir := filepath.Dir(dst)
	if err := os.MkdirAll(dir, 0o755); err != nil {
		return err
	}

	tmp, err := os.CreateTemp(dir, ".krino-*")
	if err != nil {
		return err
	}
	tmpName := tmp.Name()
	done := false
	defer func() {
		if !done {
			os.Remove(tmpName)
		}
	}()

	if _, err := io.Copy(tmp, in); err != nil {
		tmp.Close()
		return err
	}
	if err := tmp.Chmod(fi.Mode().Perm()); err != nil {
		tmp.Close()
		return err
	}
	if err := tmp.Sync(); err != nil {
		tmp.Close()
		return err
	}
	if err := tmp.Close(); err != nil {
		return err
	}
	if err := os.Chtimes(tmpName, fi.ModTime(), fi.ModTime()); err != nil {
		return err
	}

	// Deliberate, not redundant: the executor's own conflict re-check
	// (runFileStep, apply.go) already found a name free of anything on disk
	// before ever calling copyFile, so dst existing here means something
	// else claimed it in the meantime. Refusing is the only safe response —
	// silently overwriting it, via os.Rename below, would destroy whatever
	// just raced us.
	if _, err := os.Lstat(dst); err == nil {
		return fmt.Errorf("copy: destination already exists: %s", dst)
	} else if !os.IsNotExist(err) {
		return err
	}
	if err := os.Rename(tmpName, dst); err != nil {
		return err
	}
	done = true
	return nil
}

// moveFile moves src to dst. It tries os.Rename first, which is atomic when
// src and dst are on the same filesystem. Only when that fails with EXDEV
// (a different filesystem) does it fall back to copying src to dst through
// copyFile — itself leaving no temporary file and touching neither src nor
// dst on failure — and, only once that copy has landed at dst, removing src.
// A failure at any point before the copy has landed leaves src exactly
// where it was; a failure to remove src afterward leaves both a full copy
// at dst and the original at src rather than risk deleting the only good
// copy.
//
// The filesystem check is done by unwrapping the error for syscall.EXDEV,
// never by comparing a Stat_t's device field: that field's type differs
// across the platforms `make ci` vets (freebsd, openbsd), while syscall.EXDEV
// itself is defined identically on all three.
func moveFile(src, dst string) error {
	if err := os.MkdirAll(filepath.Dir(dst), 0o755); err != nil {
		return err
	}
	// This guard must hold independently of runFileStep's own pre-check,
	// layered rather than moved: trusting that some caller already checked
	// is exactly what lets a silently replacing helper cause harm. Placed
	// immediately before the operation that would otherwise clobber dst,
	// the same way copyFile's own guard sits right before its rename into
	// place.
	if err := refuseIfExists(dst); err != nil {
		return err
	}
	err := os.Rename(src, dst)
	if err == nil {
		return nil
	}
	if !errors.Is(err, syscall.EXDEV) {
		return err
	}
	if err := copyFile(src, dst); err != nil {
		return err
	}
	return os.Remove(src)
}

// renameFile renames src to dst, refusing on its own when dst already
// exists rather than trusting that a caller checked first (the same
// reasoning as moveFile's guard above): a bare os.Rename silently replaces
// an occupied destination, and runFileStep's own pre-check must not be the
// only thing standing between a rename step and that.
func renameFile(src, dst string) error {
	if err := refuseIfExists(dst); err != nil {
		return err
	}
	return os.Rename(src, dst)
}

// refuseIfExists reports an error naming dst if something is already there
// (os.Lstat succeeds, following no symlink), and propagates any other stat
// failure. A nil return means dst was confirmed absent at the moment of the
// check.
func refuseIfExists(dst string) error {
	if _, err := os.Lstat(dst); err == nil {
		return fmt.Errorf("destination already exists: %s", dst)
	} else if !os.IsNotExist(err) {
		return err
	}
	return nil
}

// maxSuffixAttempts bounds nextFreeName. internal/plan/conflict.go and
// internal/trash/trash.go each have their own cap of the same size, for the
// same reason given below: nextFreeName solves yet another, independent
// collision problem and is not sharing code with either.
const maxSuffixAttempts = 10000

// nextFreeName finds the first stem_N.ext, N starting at 1, that does not
// currently exist on disk. It is the executor's own conflict re-check (spec
// §7.4, last paragraph): planning already resolved every conflict once
// against the filesystem as it was then, but something else can claim the
// planned name before the executor gets to it, so the executor looks again,
// right before acting, using only the real filesystem — it has no run-wide
// claim set to consult, unlike planning's.
//
// internal/plan's own suffixed() and splitExt are unexported, so they are
// not reachable from here; nextFreeName and splitExt below are a second,
// small implementation, the same shape as internal/plan's and
// internal/trash's for the same reason those two do not share code with
// each other either — each resolves a distinct, independently changing set
// of collisions (planned destinations; entries already in the Trash; names
// that appeared on disk since this plan was made).
func nextFreeName(dst string) (string, error) {
	dir, base := filepath.Split(dst)
	stem, ext := splitExt(base)
	for n := 1; n <= maxSuffixAttempts; n++ {
		candidate := filepath.Join(dir, fmt.Sprintf("%s_%d%s", stem, n, ext))
		if _, err := os.Lstat(candidate); os.IsNotExist(err) {
			return candidate, nil
		}
	}
	return "", errors.New("too many conflicting names")
}

// splitExt splits name on its last dot, which does not count when it is the
// first character: "a.tar.gz" -> "a.tar", ".gz"; ".bashrc" -> ".bashrc", "".
// Duplicated from internal/plan/conflict.go and internal/trash/trash.go
// (unexported in both) rather than shared; see nextFreeName's comment.
func splitExt(name string) (stem, ext string) {
	i := strings.LastIndexByte(name, '.')
	if i <= 0 {
		return name, ""
	}
	return name[:i], name[i:]
}

// MkdirAllTracked creates dir and any missing ancestors (mode 0755),
// returning every directory it actually created, outermost first, so undo
// can later remove the empty ones again. A directory that already existed
// is not included, and nothing is created or returned on error.
func MkdirAllTracked(dir string) ([]string, error) {
	dir = filepath.Clean(dir)
	// Lstat, not Stat: a dangling symlink here reports ENOENT to Stat, so
	// the "already a directory" branch is missed, os.Mkdir then fails with
	// EEXIST, and the path used to be recorded as one krino created - which
	// undo would later remove, unlinking a symlink krino never made.
	if fi, err := os.Lstat(dir); err == nil {
		if fi.Mode()&fs.ModeSymlink != 0 {
			if resolved, serr := os.Stat(dir); serr != nil || !resolved.IsDir() {
				return nil, fmt.Errorf("%s is a symlink that does not lead to a directory", dir)
			}
			return nil, nil
		}
		if !fi.IsDir() {
			return nil, fmt.Errorf("%s exists and is not a directory", dir)
		}
		return nil, nil
	} else if !os.IsNotExist(err) {
		return nil, err
	}

	parent := filepath.Dir(dir)
	var made []string
	if parent != dir {
		parentMade, err := MkdirAllTracked(parent)
		if err != nil {
			return nil, err
		}
		made = parentMade
	}
	if err := os.Mkdir(dir, 0o755); err != nil {
		if os.IsExist(err) {
			// Something else created it between the Lstat and here: it is
			// not ours to remove again.
			return made, nil
		}
		return made, err
	}
	return append(made, dir), nil
}

// UnderSymlink reports the first component of path, at or below root, that
// is a symlink - "" when there is none, and "" when path is not under root
// at all.
//
// Placeholders are already stopped from redirecting a step out of the
// directory a rule named (internal/plan's expandDir); a symlink is a name
// too, and one planted inside the directory krino sorts - by an unpacked
// archive, say - redirects a move or a copy exactly the same way, while the
// plan the user approved shows only the text. A destination the
// configuration itself names, outside the sorted directory, is the user's
// own arrangement and is not second-guessed.
func UnderSymlink(path, root string) string {
	if root == "" {
		return ""
	}
	root = filepath.Clean(root)
	path = filepath.Clean(path)
	rel, err := filepath.Rel(root, path)
	if err != nil || rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) {
		return ""
	}
	at := root
	for _, part := range strings.Split(filepath.ToSlash(rel), "/") {
		if part == "." || part == "" {
			continue
		}
		at = filepath.Join(at, part)
		fi, err := os.Lstat(at)
		if err != nil {
			// Not there yet: nothing below it can be a symlink either.
			return ""
		}
		if fi.Mode()&fs.ModeSymlink != 0 {
			return at
		}
	}
	return ""
}