plumbing: Optimise memory consumption for filesystem storage

Previously, as part of building the index representation, the resolveObject
func would create an interim plumbing.MemoryObject, which would then be
saved into storage via storage.SetEncodedObject. This meant that objects
would be unnecessarily loaded into memory, to then be saved into disk.

The changes streamlines this process by:
- Introducing the LazyObjectWriter interface which enables the write
  operation to take places directly against the filesystem-based storage.
- Leverage multi-writers to process the input data once, while targeting
  multiple writers (e.g. hasher and storage).

An additional change relates to the caching of object info children within
Parser.get. The cache is now skipped when a seekable filesystem is being
used.

The impact of the changes can be observed when using seekable filesystem
storages, especially when cloning large repositories.

The stats below were captured by adapting the BenchmarkPlainClone test
to clone https://github.com/torvalds/linux.git:

pkg: github.com/go-git/go-git/v5
cpu: Intel(R) Core(TM) i9-10885H CPU @ 2.40GHz
              │  /tmp/old   │             /tmp/new              │
              │   sec/op    │   sec/op    vs base               │
PlainClone-16   41.68 ± 17%   48.04 ± 9%  +15.27% (p=0.015 n=6)

              │   /tmp/old    │               /tmp/new               │
              │     B/op      │     B/op       vs base               │
PlainClone-16   1127.8Mi ± 7%   256.7Mi ± 50%  -77.23% (p=0.002 n=6)

              │  /tmp/old   │              /tmp/new              │
              │  allocs/op  │  allocs/op   vs base               │
PlainClone-16   3.125M ± 0%   3.800M ± 0%  +21.60% (p=0.002 n=6)

Notice that on average the memory consumption per operation is over 75%
smaller. The time per operation increased by 15%, which may actual be less
on long running applications, due to the decreased GC pressure and the
garbage collection costs.

Signed-off-by: Paulo Gomes <pjbgf@linux.com>

1 files changed, 102 insertions, 0 deletions

diff --git a/plumbing/format/packfile/patch_delta.go b/plumbing/format/packfile/patch_delta.go
index f00562d..67c20ff 100644
--- a/plumbing/format/packfile/patch_delta.go
+++ b/plumbing/format/packfile/patch_delta.go
@@ -4,6 +4,7 @@ import (
 	"bufio"
 	"bytes"
 	"errors"
+	"fmt"
 	"io"
 	"math"
 
@@ -265,6 +266,107 @@ func patchDelta(dst *bytes.Buffer, src, delta []byte) error {
 	return nil
 }
 
+func patchDeltaWriter(dst io.Writer, base io.ReaderAt, delta io.Reader,
+	typ plumbing.ObjectType, writeHeader objectHeaderWriter) (uint, plumbing.Hash, error) {
+	deltaBuf := bufio.NewReaderSize(delta, 1024)
+	srcSz, err := decodeLEB128ByteReader(deltaBuf)
+	if err != nil {
+		if err == io.EOF {
+			return 0, plumbing.ZeroHash, ErrInvalidDelta
+		}
+		return 0, plumbing.ZeroHash, err
+	}
+
+	if r, ok := base.(*bytes.Reader); ok && srcSz != uint(r.Size()) {
+		return 0, plumbing.ZeroHash, ErrInvalidDelta
+	}
+
+	targetSz, err := decodeLEB128ByteReader(deltaBuf)
+	if err != nil {
+		if err == io.EOF {
+			return 0, plumbing.ZeroHash, ErrInvalidDelta
+		}
+		return 0, plumbing.ZeroHash, err
+	}
+
+	// If header still needs to be written, caller will provide
+	// a LazyObjectWriterHeader. This seems to be the case when
+	// dealing with thin-packs.
+	if writeHeader != nil {
+		err = writeHeader(typ, int64(targetSz))
+		if err != nil {
+			return 0, plumbing.ZeroHash, fmt.Errorf("could not lazy write header: %w", err)
+		}
+	}
+
+	remainingTargetSz := targetSz
+
+	hasher := plumbing.NewHasher(typ, int64(targetSz))
+	mw := io.MultiWriter(dst, hasher)
+
+	bufp := sync.GetByteSlice()
+	defer sync.PutByteSlice(bufp)
+
+	sr := io.NewSectionReader(base, int64(0), int64(srcSz))
+	// Keep both the io.LimitedReader types, so we can reset N.
+	baselr := io.LimitReader(sr, 0).(*io.LimitedReader)
+	deltalr := io.LimitReader(deltaBuf, 0).(*io.LimitedReader)
+
+	for {
+		buf := *bufp
+		cmd, err := deltaBuf.ReadByte()
+		if err == io.EOF {
+			return 0, plumbing.ZeroHash, ErrInvalidDelta
+		}
+		if err != nil {
+			return 0, plumbing.ZeroHash, err
+		}
+
+		if isCopyFromSrc(cmd) {
+			offset, err := decodeOffsetByteReader(cmd, deltaBuf)
+			if err != nil {
+				return 0, plumbing.ZeroHash, err
+			}
+			sz, err := decodeSizeByteReader(cmd, deltaBuf)
+			if err != nil {
+				return 0, plumbing.ZeroHash, err
+			}
+
+			if invalidSize(sz, targetSz) ||
+				invalidOffsetSize(offset, sz, srcSz) {
+				return 0, plumbing.ZeroHash, err
+			}
+
+			if _, err := sr.Seek(int64(offset), io.SeekStart); err != nil {
+				return 0, plumbing.ZeroHash, err
+			}
+			baselr.N = int64(sz)
+			if _, err := io.CopyBuffer(mw, baselr, buf); err != nil {
+				return 0, plumbing.ZeroHash, err
+			}
+			remainingTargetSz -= sz
+		} else if isCopyFromDelta(cmd) {
+			sz := uint(cmd) // cmd is the size itself
+			if invalidSize(sz, targetSz) {
+				return 0, plumbing.ZeroHash, ErrInvalidDelta
+			}
+			deltalr.N = int64(sz)
+			if _, err := io.CopyBuffer(mw, deltalr, buf); err != nil {
+				return 0, plumbing.ZeroHash, err
+			}
+
+			remainingTargetSz -= sz
+		} else {
+			return 0, plumbing.ZeroHash, err
+		}
+		if remainingTargetSz <= 0 {
+			break
+		}
+	}
+
+	return targetSz, hasher.Sum(), nil
+}
+
 // Decodes a number encoded as an unsigned LEB128 at the start of some
 // binary data and returns the decoded number and the rest of the
 // stream.