perf: reduce NewCarReader allocations

This accounted for almost 10% of the garbage on one of Bluesky's
server daemons (go tool pprof --alloc_space), leading to high GC CPU.

benchstat:

                         │   before    │                after                │
                         │   sec/op    │   sec/op     vs base                │
    NewCarReader_small-8   2.005µ ± 1%   1.451µ ± 5%  -27.61% (p=0.000 n=10)

                         │    before    │                after                 │
                         │     B/op     │     B/op      vs base                │
    NewCarReader_small-8   5.268Ki ± 0%   1.135Ki ± 0%  -78.46% (p=0.000 n=10)

                         │   before   │               after               │
                         │ allocs/op  │ allocs/op   vs base               │
    NewCarReader_small-8   29.00 ± 0%   27.00 ± 0%  -6.90% (p=0.000 n=10)

car.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"fmt"
 	"io"
+	"sync"
 
 	blocks "github.com/ipfs/go-block-format"
 	cid "github.com/ipfs/go-cid"
@@ -118,15 +119,21 @@ func (cw *carWriter) writeNode(ctx context.Context, nd format.Node) error {
 	return util.LdWrite(cw.w, nd.Cid().Bytes(), nd.RawData())
 }
 
+var bufioReaderPool = sync.Pool{
+	New: func() any { return bufio.NewReader(nil) },
+}
+
 type CarReader struct {
-	br     *bufio.Reader
+	br     *bufio.Reader // set nil on EOF
 	Header *CarHeader
 }
 
 func NewCarReader(r io.Reader) (*CarReader, error) {
-	br := bufio.NewReader(r)
+	br := bufioReaderPool.Get().(*bufio.Reader)
+	br.Reset(r)
 	ch, err := ReadHeader(br)
 	if err != nil {
+		bufioReaderPool.Put(br)
 		return nil, err
 	}
 
@@ -145,8 +152,17 @@ func NewCarReader(r io.Reader) (*CarReader, error) {
 }
 
 func (cr *CarReader) Next() (blocks.Block, error) {
+	if cr.br == nil {
+		return nil, io.EOF
+	}
 	c, data, err := util.ReadNode(cr.br)
 	if err != nil {
+		if err == io.EOF {
+			// Common happy case: recycle the bufio.Reader.
+			// In the other error paths leaking it is fine.
+			bufioReaderPool.Put(cr.br)
+			cr.br = nil
+		}
 		return nil, err
 	}
 

car_test.go

@@ -227,3 +227,27 @@ func TestBadHeaders(t *testing.T) {
 		})
 	}
 }
+
+func BenchmarkNewCarReader_small(b *testing.B) {
+	b.ReportAllocs()
+
+	fixture, err := hex.DecodeString(fixtureStr)
+	if err != nil {
+		b.Fatal(err)
+	}
+	for i := 0; i < b.N; i++ {
+		cr, err := car.NewCarReader(bytes.NewReader(fixture))
+		if err != nil {
+			b.Fatal(err)
+		}
+		for {
+			_, err := cr.Next()
+			if err == io.EOF {
+				break
+			}
+			if err != nil {
+				b.Fatal(err)
+			}
+		}
+	}
+}