perf: Micro optimization to save one allocation per packet (backport …

.changelog/unreleased/improvements/3019-reduce-allocations-in-packet-reads.md

@@ -0,0 +1,3 @@
+- [`protoio`] Remove one allocation and new object call from `ReadMsg`,
+  leading to a 4% p2p message reading performance gain.
+  ([\#3018](https://github.com/cometbft/cometbft/issues/3018)

.golangci.yml

@@ -43,6 +43,8 @@ linters-settings:
     suggest-new: true
   misspell:
     locale: US
+    ignore-words:
+      - Cancelled
   depguard:
     rules:
       main:

CHANGELOG.md

@@ -2,13 +2,15 @@
 
 ## Unreleased
 
+* [#73](https://github.com/osmosis-labs/cometbft/pull/73) perf(consensus/blockexec): Add simplistic block validation cache
+* [#74](https://github.com/osmosis-labs/cometbft/pull/74) perf(consensus): Minor speedup to mark late vote metrics
+* [#75](https://github.com/osmosis-labs/cometbft/pull/75) perf(p2p): 4% speedup to readMsg by removing one allocation
+
 ## v0.37.4-v25-osmo-3
 
 * [#61](https://github.com/osmosis-labs/cometbft/pull/61) refactor(p2p/connection): Slight refactor to sendManyPackets that helps highlight performance improvements (backport #2953) (#2978)
 * [#62](https://github.com/osmosis-labs/cometbft/pull/62) perf(consensus/blockstore): Remove validate basic call from LoadBlock
 * [#71](https://github.com/osmosis-labs/cometbft/pull/71) perf(consensus): Make mempool update async from Commit
-* [#73](https://github.com/osmosis-labs/cometbft/pull/73) perf(consensus/blockexec): Add simplistic block validation cache
-* [#74](https://github.com/osmosis-labs/cometbft/pull/74) perf(consensus): Minor speedup to mark late vote metrics
 * [#59](https://github.com/osmosis-labs/cometbft/pull/59) `[blockstore]` Remove a redundant `Header.ValidateBasic` call in `LoadBlockMeta`, 75% reducing this time. ([\#2964](https://github.com/cometbft/cometbft/pull/2964))
 * [#59](https://github.com/osmosis-labs/cometbft/pull/59) `[p2p/conn]` Speedup connection.WritePacketMsgTo, by reusing internal buffers rather than re-allocating. ([\#2986](https://github.com/cometbft/cometbft/pull/2986))
 

libs/protoio/io.go

@@ -97,3 +97,7 @@ func (r *byteReader) ReadByte() (byte, error) {
 	}
 	return r.buf[0], nil
 }
+
+func (r *byteReader) resetBytesRead() {
+	r.bytesRead = 0
+}

libs/protoio/reader.go

@@ -49,24 +49,25 @@ func NewDelimitedReader(r io.Reader, maxSize int) ReadCloser {
 	if c, ok := r.(io.Closer); ok {
 		closer = c
 	}
-	return &varintReader{r, nil, maxSize, closer}
+	return &varintReader{r, newByteReader(r), nil, maxSize, closer}
 }
 
 type varintReader struct {
-	r       io.Reader
-	buf     []byte
-	maxSize int
-	closer  io.Closer
-}
-
-func (r *varintReader) ReadMsg(msg proto.Message) (int, error) {
+	r io.Reader
 	// ReadUvarint needs an io.ByteReader, and we also need to keep track of the
 	// number of bytes read, so we use our own byteReader. This can't be
 	// buffered, so the caller should pass a buffered io.Reader to avoid poor
 	// performance.
-	byteReader := newByteReader(r.r)
-	l, err := binary.ReadUvarint(byteReader)
-	n := byteReader.bytesRead
+	byteReader *byteReader
+	buf        []byte
+	maxSize    int
+	closer     io.Closer
+}
+
+func (r *varintReader) ReadMsg(msg proto.Message) (int, error) {
+	r.byteReader.resetBytesRead()
+	l, err := binary.ReadUvarint(r.byteReader)
+	n := r.byteReader.bytesRead
 	if err != nil {
 		return n, err
 	}

libs/pubsub/subscription.go

@@ -45,8 +45,6 @@ func (s *Subscription) Out() <-chan Message {
 
 // Cancelled returns a channel that's closed when the subscription is
 // terminated and supposed to be used in a select statement.
-//
-//nolint:misspell
 func (s *Subscription) Cancelled() <-chan struct{} {
 	return s.canceled
 }

types/event_bus.go

@@ -23,7 +23,7 @@ type EventBusSubscriber interface {
 
 type Subscription interface {
 	Out() <-chan cmtpubsub.Message
-	Cancelled() <-chan struct{} //nolint: misspell
+	Cancelled() <-chan struct{}
 	Err() error
 }