runner/act/artifactcache/storage.go

// Copyright 2023 The Gitea Authors. All rights reserved.
// Copyright 2023 The nektos/act Authors. All rights reserved.
// SPDX-License-Identifier: MIT

package artifactcache

import (
	"bytes"
	"crypto/sha256"
	"encoding/hex"
	"encoding/json/v2"
	"errors"
	"fmt"
	"io"
	"net/http"
	"os"
	"path/filepath"
	"strconv"
	"strings"
)

type Storage struct {
	rootDir string
}

func NewStorage(rootDir string) (*Storage, error) {
	if err := os.MkdirAll(rootDir, 0o755); err != nil {
		return nil, err
	}
	return &Storage{
		rootDir: rootDir,
	}, nil
}

func (s *Storage) Exist(id uint64) (bool, error) {
	_, err := os.Stat(s.filename(id))
	if os.IsNotExist(err) {
		return false, nil
	}
	return err == nil, err
}

func (s *Storage) Write(id uint64, offset int64, reader io.Reader) error {
	return s.writeFile(s.tempName(id, offset), reader)
}

func (s *Storage) writeFile(name string, reader io.Reader) error {
	if err := os.MkdirAll(filepath.Dir(name), 0o755); err != nil {
		return err
	}
	file, err := os.Create(name)
	if err != nil {
		return err
	}
	_, err = io.Copy(file, reader)
	if err = errors.Join(err, file.Close()); err != nil {
		_ = os.Remove(name)
	}
	return err
}

func (s *Storage) WriteBlock(id uint64, blockID string, reader io.Reader) error {
	return s.writeFile(s.blockName(id, blockID), reader)
}

// OrderBlocks renames the staged blocks into the order the block list gives. A block the list
// does not name keeps its staged name, which is how Commit leaves it out, as Azure drops it. One
// rename pass is safe because a staged name always carries blockFilePrefix and a target name
// never does, so no rename can collide with a block not yet moved.
func (s *Storage) OrderBlocks(id uint64, blockIDs []string) error {
	wanted, err := json.Marshal(blockIDs) // a block id is client-chosen, so it cannot be a separator
	if err != nil {
		return err
	}
	list := filepath.Join(s.tempDir(id), blockFilePrefix+"list")
	if recorded, err := os.ReadFile(list); err == nil {
		// A retry repeats the list it already sent; another one would reorder what is now staged.
		if !bytes.Equal(recorded, wanted) {
			return fmt.Errorf("cache %d was already assembled from a different block list", id)
		}
		return nil
	}
	for i, blockID := range blockIDs {
		if err := os.Rename(s.blockName(id, blockID), s.tempName(id, int64(i))); err != nil {
			if os.IsNotExist(err) {
				return fmt.Errorf("block %q of cache %d was never uploaded: %w", blockID, id, err)
			}
			return err
		}
	}
	return os.WriteFile(list, wanted, 0o600)
}

func (s *Storage) Commit(id uint64, size int64) (int64, error) {
	defer func() { _ = os.RemoveAll(s.tempDir(id)) }()

	name := s.filename(id)
	tempNames, err := s.tempNames(id)
	if err != nil {
		return 0, err
	}

	if err := os.MkdirAll(filepath.Dir(name), 0o755); err != nil {
		return 0, err
	}
	written, err := assemble(name, tempNames)
	if err == nil && size >= 0 && written != size { // old clients like actions/cache@v2 send no size
		err = fmt.Errorf("broken file: %v != %v", written, size)
	}
	if err != nil {
		_ = os.Remove(name)
		return 0, err
	}
	return written, nil
}

// assemble appends the parts into the first, deleting each once copied, so it never needs twice the space.
func assemble(name string, tempNames []string) (int64, error) {
	if len(tempNames) == 0 {
		return 0, errors.New("nothing was uploaded")
	}
	if err := os.Rename(tempNames[0], name); err != nil {
		return 0, err
	}
	file, err := os.OpenFile(name, os.O_WRONLY|os.O_APPEND, 0)
	if err != nil {
		return 0, err
	}
	defer file.Close()
	for _, v := range tempNames[1:] {
		part, err := os.Open(v)
		if err != nil {
			return 0, err
		}
		_, err = io.Copy(file, part)
		_ = part.Close()
		if err != nil {
			return 0, err
		}
		_ = os.Remove(v)
	}
	info, err := file.Stat()
	if err != nil {
		return 0, err
	}
	return info.Size(), file.Close()
}

func (s *Storage) Serve(w http.ResponseWriter, r *http.Request, id uint64) {
	http.ServeFile(w, r, s.filename(id))
}

// Remove deletes an entry's blob and any staged parts. It reports failure so the caller can
// keep the entry and retry, rather than dropping the only reference to bytes on disk.
func (s *Storage) Remove(id uint64) error {
	if err := os.Remove(s.filename(id)); err != nil && !os.IsNotExist(err) {
		return err
	}
	return os.RemoveAll(s.tempDir(id))
}

func (s *Storage) filename(id uint64) string {
	return filepath.Join(s.rootDir, fmt.Sprintf("%02x", id%0xff), strconv.FormatUint(id, 10))
}

func (s *Storage) tempDir(id uint64) string {
	return filepath.Join(s.rootDir, "tmp", strconv.FormatUint(id, 10))
}

func (s *Storage) tempName(id uint64, offset int64) string {
	return filepath.Join(s.tempDir(id), fmt.Sprintf("%016x", offset))
}

// blockFilePrefix marks a staged, not yet ordered block, so that tempNames can keep it out of
// Commit's name-ordered concatenation.
const blockFilePrefix = "block-"

func (s *Storage) blockName(id uint64, blockID string) string {
	// The block id is client-chosen (base64), so it is hashed rather than trusted as a
	// path element.
	sum := sha256.Sum256([]byte(blockID))
	return filepath.Join(s.tempDir(id), blockFilePrefix+hex.EncodeToString(sum[:]))
}

func (s *Storage) tempNames(id uint64) ([]string, error) {
	dir := s.tempDir(id)
	files, err := os.ReadDir(dir)
	if err != nil {
		return nil, err
	}
	var names []string
	for _, v := range files {
		if !v.IsDir() && !strings.HasPrefix(v.Name(), blockFilePrefix) {
			names = append(names, filepath.Join(dir, v.Name()))
		}
	}
	return names, nil
}