htpack/cmd/htpacker/packer/packer.go

/*
Package packer implements the core packing functionality. It is designed to be
used by a wrapper program (CLI etc.).
*/
package packer

import (
	"bufio"
	"crypto/sha512"
	"fmt"
	"io/ioutil"
	"net/http"
	"os"
	"runtime"
	"sync"

	"github.com/andybalholm/brotli"
	"github.com/foobaz/go-zopfli/zopfli"
	"golang.org/x/sys/unix"
	"src.lwithers.me.uk/go/htpack/packed"
	"src.lwithers.me.uk/go/writefile"
)

// FilesToPack is the set of files which will be incorporated into the packfile.
// The key is the path at which the file will be served, and the value gives the
// disk filename as well as headers / options.
type FilesToPack map[string]FileToPack

// FileToPack contains the headers / options for a file which is about to be
// packed.
type FileToPack struct {
	// Filename is the path to the file on disk (relative or absolute).
	Filename string `yaml:"filename"`

	// ContentType is used as the Content-Type header for the source data.
	ContentType string `yaml:"content_type"`

	// DisableCompression can be set to skip any compression for this file.
	DisableCompression bool `yaml:"disable_compression"`

	// DisableGzip can be set to skip gzip compression for this file.
	DisableGzip bool `yaml:"disable_gzip"`

	// DisableBrotli can be set to skip brotli compression for this file.
	DisableBrotli bool `yaml:"disable_brotli"`
}

// Progress is a callback object which reports packing progress.
type Progress interface {
	// Count reports the number of items that have begun processing.
	Count(n int)

	// Begin denotes the processing of an input file.
	Begin(filename, compression string)

	// End denotes the completion of input file processing.
	End(filename, compression string)
}

type ignoreProgress int

func (ignoreProgress) Count(_ int)       {}
func (ignoreProgress) Begin(_, _ string) {}
func (ignoreProgress) End(_, _ string)   {}

const (
	// minCompressionSaving means we'll only use the compressed version of
	// the file if it's at least this many bytes smaller than the original.
	// Chosen somewhat arbitrarily; we have to add an HTTP header, and the
	// decompression overhead is not zero.
	minCompressionSaving = 128

	// minCompressionFraction means we'll only use the compressed version of
	// the file if it's at least (origSize>>minCompressionFraction) bytes
	// smaller than the original. This is a guess at when the decompression
	// overhead outweighs the time saved in transmission.
	minCompressionFraction = 7 // i.e. files must be at least 1/128 smaller

	// padWidth is the padding alignment size expressed as a power of 2.
	// The value 12 (i.e. 4096 bytes) is chosen to align with a common
	// page size and filesystem block size.
	padWidth = 12

	// sendfileLimit is the number of bytes we can transfer through a single
	// sendfile(2) call. This value is from the man page.
	sendfileLimit = 0x7FFFF000
)

// Pack a file. Use Pack2 for progress reporting.
func Pack(filesToPack FilesToPack, outputFilename string) error {
	return Pack2(filesToPack, outputFilename, nil)
}

// Pack2 will pack a file, with progress reporting. The progress interface may
// be nil.
func Pack2(filesToPack FilesToPack, outputFilename string, progress Progress) error {
	if progress == nil {
		progress = ignoreProgress(0)
	}

	finalFname, w, err := writefile.New(outputFilename)
	if err != nil {
		return err
	}
	defer writefile.Abort(w)

	// we use this little structure to serialise file writes below, and
	// it has a couple of convenience methods for common operations
	packer := packer{
		w:        w,
		progress: progress,
	}

	// write initial header (will rewrite offset/length when known)
	hdr := &packed.Header{
		Magic:           packed.Magic,
		Version:         packed.VersionInitial,
		DirectoryOffset: 1,
		DirectoryLength: 1,
	}
	m, err := hdr.Marshal()
	if err != nil {
		return fmt.Errorf("failed to marshal header (%T): %v", hdr, err)
	}
	if _, err = w.Write(m); err != nil {
		return err
	}
	if err = packer.pad(); err != nil {
		return err
	}

	// Channel to limit number of CPU-bound goroutines. One token is written
	// to the channel for each active worker; since the channel is bounded,
	// further writes will block at the limit. As workers complete, they
	// consume a token from the channel.
	nCPU := runtime.NumCPU() + 2 // +2 for I/O bound portions
	if nCPU < 4 {
		nCPU = 4
	}
	packer.cpus = make(chan struct{}, nCPU)

	// Channel to report worker errors. Writes should be non-blocking. If
	// your error is dropped, don't worry, an earlier error will be
	// reported.
	packer.errors = make(chan error, 1)

	// Channel to abort further operations. It should be closed to abort.
	// The closer should be the one who writes onto packer.errors.
	packer.aborted = make(chan struct{})

	// write the packed files, storing info for the directory structure
	packer.dir = &packed.Directory{
		Files: make(map[string]*packed.File),
	}

	var count int
PackingLoop:
	for path, fileToPack := range filesToPack {
		select {
		case <-packer.aborted:
			// a worker reported an error; break out of loop early
			break PackingLoop
		default:
			packer.packFile(path, fileToPack)
			count++
			progress.Count(count)
		}
	}

	// wait for all goroutines to complete
	for n := 0; n < nCPU; n++ {
		packer.cpus <- struct{}{}
	}

	// check whether any of the just-completed goroutines returned an error
	select {
	case err = <-packer.errors:
		return err
	default:
	}

	// write the directory
	if m, err = packer.dir.Marshal(); err != nil {
		err = fmt.Errorf("failed to marshal directory object (%T): %v",
			packer.dir, err)
		return err
	}
	dirOffset, err := w.Seek(0, os.SEEK_CUR)
	if err != nil {
		return err
	}

	if _, err := w.Write(m); err != nil {
		return err
	}

	// now modify the header at the start of the file
	hdr.DirectoryOffset = uint64(dirOffset)
	hdr.DirectoryLength = uint64(len(m))
	if m, err = hdr.Marshal(); err != nil {
		return fmt.Errorf("failed to marshal header (%T): %v", hdr, err)
	}
	if _, err = w.WriteAt(m, 0); err != nil {
		return err
	}

	// all done!
	return writefile.Commit(finalFname, w)
}

func etag(in []byte) string {
	h := sha512.New384()
	h.Write(in)
	return fmt.Sprintf(`"1--%x"`, h.Sum(nil))
}

func compressionWorthwhile(data []byte, compressed os.FileInfo) bool {
	uncompressedSize := uint64(len(data))
	sz := uint64(compressed.Size())

	switch {
	case sz+minCompressionSaving > uncompressedSize,
		sz+(uncompressedSize>>minCompressionFraction) > uncompressedSize:
		return false
	default:
		return true
	}
}

// packer packs input files into the output file. It has methods for each type
// of compression. Unexported methods assume they are called in a context where
// the lock is not needed or already taken; exported methods take the lock.
type packer struct {
	w        *os.File
	lock     sync.Mutex
	cpus     chan struct{}
	errors   chan error
	aborted  chan struct{}
	dir      *packed.Directory
	progress Progress
}

// pad will move the file write pointer to the next padding boundary. It is not
// concurrency safe.
func (p *packer) pad() error {
	pos, err := p.w.Seek(0, os.SEEK_CUR)
	if err != nil {
		return err
	}

	pos &= (1 << padWidth) - 1
	if pos == 0 { // already aligned
		return nil
	}

	_, err = p.w.Seek((1<<padWidth)-pos, os.SEEK_CUR)
	return err
}

// appendPath will copy file data from srcPath and append it to the output file. The
// offset and length are stored in ‘data’ on success. It is not concurrency safe.
func (p *packer) appendPath(srcPath string, data *packed.FileData) error {
	// open the input file and grab its length
	in, err := os.Open(srcPath)
	if err != nil {
		return err
	}
	defer in.Close()

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

	// copy in the file data
	return p.appendFile(in, fi.Size(), data)
}

// appendFile will copy file data from src and append it to the output file. The
// offset and length are stored in ‘data’ on success. It is not concurrency safe.
func (p *packer) appendFile(src *os.File, srcLen int64, data *packed.FileData) error {
	// retrieve current file position and store in data.Offset
	off, err := p.w.Seek(0, os.SEEK_CUR)
	if err != nil {
		return err
	}
	data.Length = uint64(srcLen)
	data.Offset = uint64(off)

	// copy in the file data
	remain := srcLen
	off = 0
	for remain > 0 {
		var amt int
		if remain > sendfileLimit {
			amt = sendfileLimit
		} else {
			amt = int(remain)
		}

		amt, err := unix.Sendfile(int(p.w.Fd()), int(src.Fd()), &off, amt)
		remain -= int64(amt)
		if err != nil {
			return fmt.Errorf("sendfile (copying data to "+
				"htpack): %v", err)
		}
	}

	// leave output file padded to next boundary
	return p.pad()
}

func (p *packer) packFile(path string, fileToPack FileToPack) {
	// open and mmap input file
	f, err := os.Open(fileToPack.Filename)
	if err != nil {
		p.Abort(err)
		return
	}
	defer f.Close()

	fi, err := f.Stat()
	if err != nil {
		p.Abort(err)
		return
	}

	var data []byte
	if fi.Size() > 0 {
		data, err = unix.Mmap(int(f.Fd()), 0, int(fi.Size()),
			unix.PROT_READ, unix.MAP_SHARED)
		if err != nil {
			p.Abort(fmt.Errorf("mmap %s: %v", fileToPack.Filename, err))
			return
		}
	}

	// prepare initial directory entry
	info := &packed.File{
		Etag:        etag(data),
		ContentType: fileToPack.ContentType,
	}
	if info.ContentType == "" {
		info.ContentType = http.DetectContentType(data)
	}
	p.dir.Files[path] = info // NB: this part is not concurrent, so no mutex

	// list of operations on this input file that we'll carry out asynchronously
	ops := []func() error{
		func() error {
			p.progress.Begin(fileToPack.Filename, "uncompressed")
			defer p.progress.End(fileToPack.Filename, "uncompressed")
			return p.Uncompressed(fileToPack.Filename, info)
		},
	}
	if !fileToPack.DisableCompression && !fileToPack.DisableGzip {
		ops = append(ops, func() error {
			p.progress.Begin(fileToPack.Filename, "gzip")
			defer p.progress.End(fileToPack.Filename, "gzip")
			if err := p.Gzip(data, info); err != nil {
				return fmt.Errorf("gzip compression of %s "+
					"failed: %v", fileToPack.Filename, err)
			}
			return nil
		})
	}
	if !fileToPack.DisableCompression && !fileToPack.DisableBrotli {
		ops = append(ops, func() error {
			p.progress.Begin(fileToPack.Filename, "brotli")
			defer p.progress.End(fileToPack.Filename, "brotli")
			if err := p.Brotli(data, info); err != nil {
				return fmt.Errorf("brotli compression of %s "+
					"failed: %v", fileToPack.Filename, err)
			}
			return nil
		})
	}

	// we have multiple operations on the file, and we need to wait for
	// them all to complete before munmap()
	wg := new(sync.WaitGroup)
	wg.Add(len(ops))
	go func() {
		wg.Wait()
		unix.Munmap(data)
	}()

	for _, op := range ops {
		select {
		case <-p.aborted:
			// skip the operation
			wg.Done()

		case p.cpus <- struct{}{}:
			go func(op func() error) {
				if err := op(); err != nil {
					p.Abort(err)
				}
				// release CPU token
				<-p.cpus
				wg.Done()
			}(op)
		}
	}

	return
}

// Abort records that an error occurred and records it onto the errors channel.
// It signals workers to abort by closed the aborted channel. If called
// multiple times, only one error will be recorded, and the aborted channel will
// only be closed once.
func (p *packer) Abort(err error) {
	select {
	case p.errors <- err:
		// only one error can be written to this channel, so the write
		// acts as a lock to ensure only a single close operation takes
		// place
		close(p.aborted)
	default:
		// errors channel was already written, so we're already aborted
	}
}

// Uncompressed copies in an uncompressed file.
func (p *packer) Uncompressed(srcPath string, dir *packed.File) error {
	dir.Uncompressed = new(packed.FileData)
	p.lock.Lock()
	defer p.lock.Unlock()
	return p.appendPath(srcPath, dir.Uncompressed)
}

// Gzip will gzip input data to a temporary file, and then append that to the
// output file.
func (p *packer) Gzip(data []byte, dir *packed.File) error {
	// write via temporary file
	tmpfile, err := ioutil.TempFile("", "")
	if err != nil {
		return err
	}
	defer os.Remove(tmpfile.Name())
	defer tmpfile.Close()

	// compress
	opts := zopfli.DefaultOptions()
	if len(data) > (10 << 20) { // 10MiB
		opts.NumIterations = 5
	}

	buf := bufio.NewWriter(tmpfile)
	if err = zopfli.GzipCompress(&opts, data, buf); err != nil {
		return err
	}
	if err = buf.Flush(); err != nil {
		return err
	}

	// grab file length, evaluate whether compression is worth it
	fi, err := tmpfile.Stat()
	if err != nil {
		return err
	}
	if !compressionWorthwhile(data, fi) {
		return nil
	}

	// save the compressed data
	dir.Gzip = new(packed.FileData)
	p.lock.Lock()
	defer p.lock.Unlock()
	return p.appendFile(tmpfile, fi.Size(), dir.Gzip)
}

// Brotli will compress input data to a temporary file, and then append that to
// the output file.
func (p *packer) Brotli(data []byte, dir *packed.File) error {
	// write via temporary file
	tmpfile, err := ioutil.TempFile("", "")
	if err != nil {
		return err
	}
	defer os.Remove(tmpfile.Name())
	defer tmpfile.Close()

	// compress
	buf := bufio.NewWriter(tmpfile)
	comp := brotli.NewWriterOptions(buf, brotli.WriterOptions{
		Quality: 11,
	})
	if _, err = comp.Write(data); err != nil {
		return err
	}
	if err = comp.Close(); err != nil {
		return err
	}
	if err = buf.Flush(); err != nil {
		return err
	}

	// grab file length, evaluate whether compression is worth it
	fi, err := tmpfile.Stat()
	if err != nil {
		return err
	}
	if !compressionWorthwhile(data, fi) {
		return nil
	}

	// save the compressed data
	dir.Brotli = new(packed.FileData)
	p.lock.Lock()
	defer p.lock.Unlock()
	return p.appendFile(tmpfile, fi.Size(), dir.Brotli)
}