plumbing: format/packfile, performance optimizations for reading large commit histories (#963)

Signed-off-by: Filip Navara <navara@emclient.com>

5 files changed, 89 insertions, 33 deletions

diff --git a/plumbing/format/packfile/fsobject.go b/plumbing/format/packfile/fsobject.go
index 330cb73..a268bce 100644
--- a/plumbing/format/packfile/fsobject.go
+++ b/plumbing/format/packfile/fsobject.go
@@ -48,7 +48,7 @@ func NewFSObject(
 // Reader implements the plumbing.EncodedObject interface.
 func (o *FSObject) Reader() (io.ReadCloser, error) {
 	obj, ok := o.cache.Get(o.hash)
-	if ok {
+	if ok && obj != o {
 		reader, err := obj.Reader()
 		if err != nil {
 			return nil, err
diff --git a/plumbing/format/packfile/packfile.go b/plumbing/format/packfile/packfile.go
index 2166e0a..1e7ef26 100644
--- a/plumbing/format/packfile/packfile.go
+++ b/plumbing/format/packfile/packfile.go
@@ -21,6 +21,16 @@ var (
 	ErrZLib = NewError("zlib reading error")
 )
 
+// When reading small objects from packfile it is beneficial to do so at
+// once to exploit the buffered I/O. In many cases the objects are so small
+// that they were already loaded to memory when the object header was
+// loaded from the packfile. Wrapping in FSObject would cause this buffered
+// data to be thrown away and then re-read later, with the additional
+// seeking causing reloads from disk. Objects smaller than this threshold
+// are now always read into memory and stored in cache instead of being
+// wrapped in FSObject.
+const smallObjectThreshold = 16 * 1024
+
 // Packfile allows retrieving information from inside a packfile.
 type Packfile struct {
 	idxfile.Index
@@ -79,15 +89,7 @@ func (p *Packfile) GetByOffset(o int64) (plumbing.EncodedObject, error) {
 		}
 	}
 
-	if _, err := p.s.SeekFromStart(o); err != nil {
-		if err == io.EOF || isInvalid(err) {
-			return nil, plumbing.ErrObjectNotFound
-		}
-
-		return nil, err
-	}
-
-	return p.nextObject()
+	return p.objectAtOffset(o)
 }
 
 // GetSizeByOffset retrieves the size of the encoded object from the
@@ -108,6 +110,12 @@ func (p *Packfile) GetSizeByOffset(o int64) (size int64, err error) {
 	return h.Length, nil
 }
 
+func (p *Packfile) objectHeaderAtOffset(offset int64) (*ObjectHeader, error) {
+	h, err := p.s.SeekObjectHeader(offset)
+	p.s.pendingObject = nil
+	return h, err
+}
+
 func (p *Packfile) nextObjectHeader() (*ObjectHeader, error) {
 	h, err := p.s.NextObjectHeader()
 	p.s.pendingObject = nil
@@ -154,11 +162,7 @@ func (p *Packfile) getObjectType(h *ObjectHeader) (typ plumbing.ObjectType, err
 		if baseType, ok := p.offsetToType[offset]; ok {
 			typ = baseType
 		} else {
-			if _, err = p.s.SeekFromStart(offset); err != nil {
-				return
-			}
-
-			h, err = p.nextObjectHeader()
+			h, err = p.objectHeaderAtOffset(offset)
 			if err != nil {
 				return
 			}
@@ -175,8 +179,8 @@ func (p *Packfile) getObjectType(h *ObjectHeader) (typ plumbing.ObjectType, err
 	return
 }
 
-func (p *Packfile) nextObject() (plumbing.EncodedObject, error) {
-	h, err := p.nextObjectHeader()
+func (p *Packfile) objectAtOffset(offset int64) (plumbing.EncodedObject, error) {
+	h, err := p.objectHeaderAtOffset(offset)
 	if err != nil {
 		if err == io.EOF || isInvalid(err) {
 			return nil, plumbing.ErrObjectNotFound
@@ -190,6 +194,13 @@ func (p *Packfile) nextObject() (plumbing.EncodedObject, error) {
 		return p.getNextObject(h)
 	}
 
+	// If the object is not a delta and it's small enough then read it
+	// completely into memory now since it is already read from disk
+	// into buffer anyway.
+	if h.Length <= smallObjectThreshold && h.Type != plumbing.OFSDeltaObject && h.Type != plumbing.REFDeltaObject {
+		return p.getNextObject(h)
+	}
+
 	hash, err := p.FindHash(h.Offset)
 	if err != nil {
 		return nil, err
@@ -233,11 +244,7 @@ func (p *Packfile) getObjectContent(offset int64) (io.ReadCloser, error) {
 		}
 	}
 
-	if _, err := p.s.SeekFromStart(offset); err != nil {
-		return nil, err
-	}
-
-	h, err := p.nextObjectHeader()
+	h, err := p.objectHeaderAtOffset(offset)
 	if err != nil {
 		return nil, err
 	}
@@ -329,8 +336,6 @@ func (p *Packfile) fillOFSDeltaObjectContent(obj plumbing.EncodedObject, offset
 		if err != nil {
 			return err
 		}
-
-		p.cachePut(base)
 	}
 
 	obj.SetType(base.Type())
diff --git a/plumbing/format/packfile/parser.go b/plumbing/format/packfile/parser.go
index 5a62d63..71cbba9 100644
--- a/plumbing/format/packfile/parser.go
+++ b/plumbing/format/packfile/parser.go
@@ -398,11 +398,7 @@ func (p *Parser) readData(o *objectInfo) ([]byte, error) {
 		return data, nil
 	}
 
-	if _, err := p.scanner.SeekFromStart(o.Offset); err != nil {
-		return nil, err
-	}
-
-	if _, err := p.scanner.NextObjectHeader(); err != nil {
+	if _, err := p.scanner.SeekObjectHeader(o.Offset); err != nil {
 		return nil, err
 	}
 
diff --git a/plumbing/format/packfile/scanner.go b/plumbing/format/packfile/scanner.go
index 6fc183b..614b0d1 100644
--- a/plumbing/format/packfile/scanner.go
+++ b/plumbing/format/packfile/scanner.go
@@ -138,14 +138,52 @@ func (s *Scanner) readCount() (uint32, error) {
 	return binary.ReadUint32(s.r)
 }
 
+// SeekObjectHeader seeks to specified offset and returns the ObjectHeader
+// for the next object in the reader
+func (s *Scanner) SeekObjectHeader(offset int64) (*ObjectHeader, error) {
+	// if seeking we assume that you are not interested in the header
+	if s.version == 0 {
+		s.version = VersionSupported
+	}
+
+	if _, err := s.r.Seek(offset, io.SeekStart); err != nil {
+		return nil, err
+	}
+
+	h, err := s.nextObjectHeader()
+	if err != nil {
+		return nil, err
+	}
+
+	h.Offset = offset
+	return h, nil
+}
+
 // NextObjectHeader returns the ObjectHeader for the next object in the reader
 func (s *Scanner) NextObjectHeader() (*ObjectHeader, error) {
-	defer s.Flush()
-
 	if err := s.doPending(); err != nil {
 		return nil, err
 	}
 
+	offset, err := s.r.Seek(0, io.SeekCurrent)
+	if err != nil {
+		return nil, err
+	}
+
+	h, err := s.nextObjectHeader()
+	if err != nil {
+		return nil, err
+	}
+
+	h.Offset = offset
+	return h, nil
+}
+
+// nextObjectHeader returns the ObjectHeader for the next object in the reader
+// without the Offset field
+func (s *Scanner) nextObjectHeader() (*ObjectHeader, error) {
+	defer s.Flush()
+
 	s.crc.Reset()
 
 	h := &ObjectHeader{}
@@ -308,7 +346,7 @@ var byteSlicePool = sync.Pool{
 // SeekFromStart sets a new offset from start, returns the old position before
 // the change.
 func (s *Scanner) SeekFromStart(offset int64) (previous int64, err error) {
-	// if seeking we assume that you are not interested on the header
+	// if seeking we assume that you are not interested in the header
 	if s.version == 0 {
 		s.version = VersionSupported
 	}
@@ -385,7 +423,7 @@ type bufferedSeeker struct {
 }
 
 func (r *bufferedSeeker) Seek(offset int64, whence int) (int64, error) {
-	if whence == io.SeekCurrent {
+	if whence == io.SeekCurrent && offset == 0 {
 		current, err := r.r.Seek(offset, whence)
 		if err != nil {
 			return current, err
diff --git a/plumbing/format/packfile/scanner_test.go b/plumbing/format/packfile/scanner_test.go
index 644d0eb..091b457 100644
--- a/plumbing/format/packfile/scanner_test.go
+++ b/plumbing/format/packfile/scanner_test.go
@@ -118,6 +118,23 @@ func (s *ScannerSuite) TestNextObjectHeaderWithOutReadObjectNonSeekable(c *C) {
 	c.Assert(n, Equals, f.PackfileHash)
 }
 
+func (s *ScannerSuite) TestSeekObjectHeader(c *C) {
+	r := fixtures.Basic().One().Packfile()
+	p := NewScanner(r)
+
+	h, err := p.SeekObjectHeader(expectedHeadersOFS[4].Offset)
+	c.Assert(err, IsNil)
+	c.Assert(h, DeepEquals, &expectedHeadersOFS[4])
+}
+
+func (s *ScannerSuite) TestSeekObjectHeaderNonSeekable(c *C) {
+	r := io.MultiReader(fixtures.Basic().One().Packfile())
+	p := NewScanner(r)
+
+	_, err := p.SeekObjectHeader(expectedHeadersOFS[4].Offset)
+	c.Assert(err, Equals, ErrSeekNotSupported)
+}
+
 var expectedHeadersOFS = []ObjectHeader{
 	{Type: plumbing.CommitObject, Offset: 12, Length: 254},
 	{Type: plumbing.OFSDeltaObject, Offset: 186, Length: 93, OffsetReference: 12},