imgproxy/asyncbuffer/buffer.go

// Package asyncbuffer provides an asynchronous buffer that reads data from an
// io.Reader in the background.
//
// When created, AsyncBuffer starts reading from the upstream reader in the
// background. If a read error occurs, it is stored and can be checked with
// AsyncBuffer.Error().
//
// When reading through AsyncBuffer.Reader().Read(), the error is only returned
// once the reader reaches the point where the error occurred. In other words,
// errors are delayed until encountered by the reader.
//
// However, AsyncBuffer.Close() and AsyncBuffer.Error() will immediately return
// any stored error, even if the reader has not yet reached the error point.
package asyncbuffer

import (
	"context"
	"errors"
	"io"
	"log/slog"
	"sync"
	"sync/atomic"

	"github.com/imgproxy/imgproxy/v3/ierrors"
	"github.com/imgproxy/imgproxy/v3/ioutil"
)

const (
	// chunkSize is the size of each chunk in bytes
	chunkSize = 4096

	// pauseThreshold is the size of the file which is always read to memory. Data beyond the
	// threshold is read only if accessed. If not a multiple of chunkSize, the last chunk it points
	// to is read in full.
	pauseThreshold = 32768 // 32 KiB
)

// byteChunk is a struct that holds a buffer and the data read from the upstream reader
// data slice is required since the chunk read may be smaller than ChunkSize
type byteChunk struct {
	buf  []byte
	data []byte
}

// chunkPool is a global sync.Pool that holds byteChunk objects for
// all readers
var chunkPool = sync.Pool{
	New: func() any {
		buf := make([]byte, chunkSize)

		return &byteChunk{
			buf:  buf,
			data: buf[:0],
		}
	},
}

// AsyncBuffer is a wrapper around io.Reader that reads data in chunks
// in background and allows reading from synchronously.
type AsyncBuffer struct {
	r       io.ReadCloser // Upstream reader
	dataLen int           // Expected length of the data in r, <= 0 means unknown length

	chunks []*byteChunk // References to the chunks read from the upstream reader
	mu     sync.RWMutex // Mutex on chunks slice

	err       atomic.Value // Error that occurred during reading
	bytesRead atomic.Int64 // Total length of the data read

	finished atomic.Bool // Indicates that the buffer has finished reading
	closed   atomic.Bool // Indicates that the buffer was closed

	paused    *Latch // Paused buffer does not read data beyond threshold
	chunkCond *Cond  // Ticker that signals when a new chunk is ready

	finishOnce sync.Once
	finishFn   []context.CancelFunc
}

// New creates a new AsyncBuffer that reads from the given io.ReadCloser in background
// and closes it when finished.
//
//	r - io.ReadCloser to read data from
//	dataLen - expected length of the data in r, <= 0 means unknown length
//	finishFn - optional functions to call when the buffer is finished reading
func New(r io.ReadCloser, dataLen int, finishFn ...context.CancelFunc) *AsyncBuffer {
	ab := &AsyncBuffer{
		r:         r,
		dataLen:   dataLen,
		paused:    NewLatch(),
		chunkCond: NewCond(),
		finishFn:  finishFn,
	}

	go ab.readChunks()

	return ab
}

// NewReadFull creates a new AsyncBuffer that reads from the given io.ReadCloser
// in foreground, blocking until all data is read. It returns an error if reading
// fails. When read fails, the reader is closed and resources are released immediately.
func NewReadFull(r io.ReadCloser, dataLen int, finishFn ...context.CancelFunc) (*AsyncBuffer, error) {
	ab := &AsyncBuffer{
		r:         r,
		dataLen:   dataLen,
		paused:    NewLatch(),
		chunkCond: NewCond(),
		finishFn:  finishFn,
	}

	// Release the paused latch so that the reader can read all data immediately
	ab.paused.Release()

	// Read all data in foreground
	ab.readChunks()

	// If error occurred during reading, return it
	if ab.Error() != nil {
		ab.Close() // Reader should be closed and resources released
		return nil, ab.Error()
	}

	return ab, nil
}

// callFinishFn calls the finish functions registered with the AsyncBuffer.
func (ab *AsyncBuffer) callFinishFn() {
	ab.finishOnce.Do(func() {
		for _, fn := range ab.finishFn {
			if fn != nil {
				fn()
			}
		}
	})
}

func (ab *AsyncBuffer) setErr(err error) {
	if err == nil {
		return
	}

	// If the error is already set, we do not overwrite it
	if ab.err.Load() == nil {
		ab.err.Store(ierrors.Wrap(err, 1))
	}
}

// addChunk adds a new chunk to the AsyncBuffer, increments bytesRead
// and signals that a chunk is ready
func (ab *AsyncBuffer) addChunk(chunk *byteChunk) {
	ab.mu.Lock()
	defer ab.mu.Unlock()

	if ab.closed.Load() {
		// If the reader is closed, we return the chunk to the pool
		chunkPool.Put(chunk)
		return
	}

	// Store the chunk, increase chunk size, increase length of the data read
	ab.chunks = append(ab.chunks, chunk)
	ab.bytesRead.Add(int64(len(chunk.data)))

	ab.chunkCond.Tick()
}

// readChunks reads data from the upstream reader in background and stores them in the pool
func (ab *AsyncBuffer) readChunks() {
	defer func() {
		if ab.bytesRead.Load() < int64(ab.dataLen) {
			// If the reader has finished reading and we have not read enough data,
			// set err to io.ErrUnexpectedEOF
			ab.setErr(io.ErrUnexpectedEOF)
		}

		// Indicate that the reader has finished reading
		ab.finished.Store(true)
		ab.chunkCond.Close()

		// Close the upstream reader
		if err := ab.r.Close(); err != nil {
			slog.Warn(
				"error closing upstream reader",
				"error", err,
				"source", "asyncbuffer.AsyncBuffer.readChunks",
			)
		}

		ab.callFinishFn()
	}()

	r := ab.r.(io.Reader)
	if ab.dataLen > 0 {
		// If the data length is known, we read only that much data
		r = io.LimitReader(r, int64(ab.dataLen))
	}

	// Stop reading if the reader is closed
	for !ab.closed.Load() {
		// In case we are trying to read data beyond threshold and we are paused,
		// wait for pause to be released.
		if ab.bytesRead.Load() >= pauseThreshold {
			ab.paused.Wait()

			// If the reader has been closed while waiting, we can stop reading
			if ab.closed.Load() {
				return // No more data to read
			}
		}

		// Get a chunk from the pool
		// If the pool is empty, it will create a new byteChunk with ChunkSize
		chunk, ok := chunkPool.Get().(*byteChunk)
		if !ok {
			ab.setErr(errors.New("asyncbuffer.AsyncBuffer.readChunks: failed to get chunk from pool"))
			return
		}

		// Read data into the chunk's buffer
		// There is no way to guarantee that r.Read will abort on context cancellation,
		// unfortunately, this is how golang works.
		n, err := ioutil.TryReadFull(r, chunk.buf)

		// If it's not the EOF, we need to store the error
		if err != nil && err != io.EOF {
			ab.setErr(err)
			chunkPool.Put(chunk)
			return
		}

		// No bytes were read (n == 0), we can return the chunk to the pool
		if n == 0 {
			chunkPool.Put(chunk)
			return
		}

		// Resize the chunk's data slice to the number of bytes read
		chunk.data = chunk.buf[:n]

		// Store the reference to the chunk in the AsyncBuffer
		ab.addChunk(chunk)

		// EOF at this point means that some bytes were read, but this is the
		// end of the stream, so we can stop reading
		if err == io.EOF {
			return
		}
	}
}

// closedError returns an error if the attempt to read on a closed reader was made.
// If the reader had an error, it returns that error instead.
func (ab *AsyncBuffer) closedError() error {
	// If the reader is closed, we return the error or nil
	if !ab.closed.Load() {
		return nil
	}

	err := ab.Error()
	if err == nil {
		err = errors.New("asyncbuffer.AsyncBuffer.ReadAt: attempt to read on closed reader")
	}

	return err
}

// offsetAvailable checks if the data at the given offset is available for reading.
// It may return io.EOF if the reader is finished reading and the offset is beyond the end of the stream.
func (ab *AsyncBuffer) offsetAvailable(off int64) (bool, error) {
	// We can not read data from the closed reader, none
	if err := ab.closedError(); err != nil {
		return false, err
	}

	// In case the offset falls within the already read chunks, we can return immediately,
	// even if error has occurred in the future
	if off < ab.bytesRead.Load() {
		return true, nil
	}

	// In case the reader is finished reading, and we have not read enough
	// data yet, return either error or EOF
	if ab.finished.Load() {
		// In case, error has occurred, we need to return it
		if err := ab.Error(); err != nil {
			return false, err
		}

		// Otherwise, it's EOF if the offset is beyond the end of the stream
		return false, io.EOF
	}

	// No available data
	return false, nil
}

// WaitFor waits for the data to be ready at the given offset. nil means ok.
// It guarantees that the chunk at the given offset is ready to be read.
func (ab *AsyncBuffer) WaitFor(off int64) error {
	// In case we are trying to read data which would potentially hit the pause threshold,
	// we need to unpause the reader ASAP.
	if off >= pauseThreshold {
		ab.paused.Release()
	}

	for {
		ok, err := ab.offsetAvailable(off)
		if ok || err != nil {
			return err
		}

		ab.chunkCond.Wait()
	}
}

// Wait waits for the reader to finish reading all data and returns
// the total length of the data read.
func (ab *AsyncBuffer) Wait() (int, error) {
	// Wait ends till the end of the stream: unpause the reader
	ab.paused.Release()

	for {
		// We can not read data from the closed reader
		if err := ab.closedError(); err != nil {
			return 0, err
		}

		// In case the reader is finished reading, we can return immediately
		if ab.finished.Load() {
			return int(ab.bytesRead.Load()), ab.Error()
		}

		// Lock until the next chunk is ready
		ab.chunkCond.Wait()
	}
}

// ReleaseThreshold releases the pause, allowing the buffer to immediately
// read data beyond the pause threshold.
func (ab *AsyncBuffer) ReleaseThreshold() {
	ab.paused.Release()
}

// Error returns the error that occurred during reading data in background.
func (ab *AsyncBuffer) Error() error {
	err := ab.err.Load()
	if err == nil {
		return nil
	}

	errCast, ok := err.(error)
	if !ok {
		return errors.New("asyncbuffer.AsyncBuffer.Error: failed to get error")
	}

	return errCast
}

// readChunkAt copies data from the chunk at the given absolute offset to the provided slice.
// Chunk must be available when this method is called.
// Returns the number of bytes copied to the slice or 0 if chunk has no data
// (eg. offset is beyond the end of the stream).
func (ab *AsyncBuffer) readChunkAt(p []byte, off int64) int {
	// If the chunk is not available, we return 0
	if off >= ab.bytesRead.Load() {
		return 0
	}

	ind := off / chunkSize // chunk index
	chunk := ab.chunks[ind]

	startOffset := off % chunkSize // starting offset in the chunk

	// If the offset in current chunk is greater than the data
	// it has, we return 0
	if startOffset >= int64(len(chunk.data)) {
		return 0
	}

	// Copy data to the target slice. The number of bytes to copy is limited by the
	// size of the target slice and the size of the data in the chunk.
	return copy(p, chunk.data[startOffset:])
}

// readAt reads data from the AsyncBuffer at the given offset.
//
// Please note that if pause threshold is hit in the middle of the reading,
// the data beyond the threshold may not be available.
//
// If the reader is paused and we try to read data beyond the pause threshold,
// it will wait till something could be returned.
func (ab *AsyncBuffer) readAt(p []byte, off int64) (int, error) {
	size := int64(len(p)) // total size of the data to read

	if off < 0 {
		return 0, errors.New("asyncbuffer.AsyncBuffer.readAt: negative offset")
	}

	// If we plan to hit threshold while reading, release the paused reader
	if int64(len(p))+off > pauseThreshold {
		ab.paused.Release()
	}

	// Wait for the offset to be available.
	// It may return io.EOF if the offset is beyond the end of the stream.
	err := ab.WaitFor(off)
	if err != nil {
		return 0, err
	}

	// We lock the mutex until current buffer is read
	ab.mu.RLock()
	defer ab.mu.RUnlock()

	// If the reader is closed, we return an error
	if err := ab.closedError(); err != nil {
		return 0, err
	}

	// Read data from the first chunk
	n := ab.readChunkAt(p, off)
	if n == 0 {
		return 0, io.EOF // Failed to read any data: means we tried to read beyond the end of the stream
	}

	size -= int64(n)
	off += int64(n) // Here and beyond off always points to the last read byte + 1

	// Now, let's try to read the rest of the data from next chunks while they are available
	for size > 0 {
		// If data is not available at the given offset, we can return data read so far.
		ok, err := ab.offsetAvailable(off)
		if !ok {
			if err == io.EOF {
				return n, nil
			}

			return n, err
		}

		// Read data from the next chunk
		nX := ab.readChunkAt(p[n:], off)
		n += nX
		size -= int64(nX)
		off += int64(nX)

		// If we read data shorter than ChunkSize or, in case that was the last chunk, less than
		// the size of the tail, return kind of EOF
		if int64(nX) < min(size, int64(chunkSize)) {
			return n, nil
		}
	}

	return n, nil
}

// Close closes the AsyncBuffer and releases all resources. It is idempotent.
func (ab *AsyncBuffer) Close() error {
	ab.mu.Lock()
	defer ab.mu.Unlock()

	// If the reader is already closed, we return immediately error or nil
	if ab.closed.Load() {
		return nil
	}

	ab.closed.Store(true)

	// Return all chunks to the pool
	for _, chunk := range ab.chunks {
		chunkPool.Put(chunk)
	}

	// Release the paused latch so that no goroutines are waiting for it
	ab.paused.Release()

	// Finish downloading
	ab.callFinishFn()

	return nil
}

// Reader returns an io.ReadSeeker+io.ReaderAt that can be used to read actual data from the AsyncBuffer
func (ab *AsyncBuffer) Reader() *Reader {
	return &Reader{ab: ab, pos: 0}
}