Commit 10c134ea authored by Tom Bergan's avatar Tom Bergan

http2: replace fixedBuffer with dataBuffer

fixedBuffer was a bad idea for two reasons:

1. It was fixed at a constant 64KB (the current default flow-control
   window) which wastes memory on the server when clients upload many
   small request bodies.

2. A follow-up CL will allow configuring the server's connection and
   stream receive windows. We want to allow individual streams to use
   varying amounts of the available connection window. This is not
   possible when each stream uses a fixedBuffer.

dataBuffer grows and shrinks based on current usage. The worst-case
fragmentation of dataBuffer is 32KB wasted memory per stream, but I
expect that worst-case will be rare. In particular, if the declared
size of a stream's request body is under 1KB, then the server will not
allocate more than 1KB to process that stream's request body.

Updates golang/go#16512
Fixes golang/go#18509

Change-Id: Ibcb18007037e82518a65848ef3baf4937955ac9d
Reviewed-on: https://go-review.googlesource.com/37400
Run-TryBot: Tom Bergan <tombergan@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: 's avatarBrad Fitzpatrick <bradfitz@golang.org>
parent dd2d9a67
// Copyright 2014 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package http2
import (
"errors"
"fmt"
"sync"
)
// Buffer chunks are allocated from a pool to reduce pressure on GC.
// The maximum wasted space per dataBuffer is 2x the largest size class,
// which happens when the dataBuffer has multiple chunks and there is
// one unread byte in both the first and last chunks. We use a few size
// classes to minimize overheads for servers that typically receive very
// small request bodies.
//
// TODO: Benchmark to determine if the pools are necessary. The GC may have
// improved enough that we can instead allocate chunks like this:
// make([]byte, max(16<<10, expectedBytesRemaining))
var (
dataChunkSizeClasses = []int{
1 << 10,
2 << 10,
4 << 10,
8 << 10,
16 << 10,
}
dataChunkPools = [...]sync.Pool{
{New: func() interface{} { return make([]byte, 1<<10) }},
{New: func() interface{} { return make([]byte, 2<<10) }},
{New: func() interface{} { return make([]byte, 4<<10) }},
{New: func() interface{} { return make([]byte, 8<<10) }},
{New: func() interface{} { return make([]byte, 16<<10) }},
}
)
func getDataBufferChunk(size int64) []byte {
i := 0
for ; i < len(dataChunkSizeClasses)-1; i++ {
if size <= int64(dataChunkSizeClasses[i]) {
break
}
}
return dataChunkPools[i].Get().([]byte)
}
func putDataBufferChunk(p []byte) {
for i, n := range dataChunkSizeClasses {
if len(p) == n {
dataChunkPools[i].Put(p)
return
}
}
panic(fmt.Sprintf("unexpected buffer len=%v", len(p)))
}
// dataBuffer is an io.ReadWriter backed by a list of data chunks.
// Each dataBuffer is used to read DATA frames on a single stream.
// The buffer is divided into chunks so the server can limit the
// total memory used by a single connection without limiting the
// request body size on any single stream.
type dataBuffer struct {
chunks [][]byte
r int // next byte to read is chunks[0][r]
w int // next byte to write is chunks[len(chunks)-1][w]
size int // total buffered bytes
expected int64 // we expect at least this many bytes in future Write calls (ignored if <= 0)
}
var errReadEmpty = errors.New("read from empty dataBuffer")
// Read copies bytes from the buffer into p.
// It is an error to read when no data is available.
func (b *dataBuffer) Read(p []byte) (int, error) {
if b.size == 0 {
return 0, errReadEmpty
}
var ntotal int
for len(p) > 0 && b.size > 0 {
readFrom := b.bytesFromFirstChunk()
n := copy(p, readFrom)
p = p[n:]
ntotal += n
b.r += n
b.size -= n
// If the first chunk has been consumed, advance to the next chunk.
if b.r == len(b.chunks[0]) {
putDataBufferChunk(b.chunks[0])
end := len(b.chunks) - 1
copy(b.chunks[:end], b.chunks[1:])
b.chunks[end] = nil
b.chunks = b.chunks[:end]
b.r = 0
}
}
return ntotal, nil
}
func (b *dataBuffer) bytesFromFirstChunk() []byte {
if len(b.chunks) == 1 {
return b.chunks[0][b.r:b.w]
}
return b.chunks[0][b.r:]
}
// Len returns the number of bytes of the unread portion of the buffer.
func (b *dataBuffer) Len() int {
return b.size
}
// Write appends p to the buffer.
func (b *dataBuffer) Write(p []byte) (int, error) {
ntotal := len(p)
for len(p) > 0 {
// If the last chunk is empty, allocate a new chunk. Try to allocate
// enough to fully copy p plus any additional bytes we expect to
// receive. However, this may allocate less than len(p).
want := int64(len(p))
if b.expected > want {
want = b.expected
}
chunk := b.lastChunkOrAlloc(want)
n := copy(chunk[b.w:], p)
p = p[n:]
b.w += n
b.size += n
b.expected -= int64(n)
}
return ntotal, nil
}
func (b *dataBuffer) lastChunkOrAlloc(want int64) []byte {
if len(b.chunks) != 0 {
last := b.chunks[len(b.chunks)-1]
if b.w < len(last) {
return last
}
}
chunk := getDataBufferChunk(want)
b.chunks = append(b.chunks, chunk)
b.w = 0
return chunk
}
// Copyright 2017 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package http2
import (
"bytes"
"fmt"
"reflect"
"testing"
)
func fmtDataChunk(chunk []byte) string {
out := ""
var last byte
var count int
for _, c := range chunk {
if c != last {
if count > 0 {
out += fmt.Sprintf(" x %d ", count)
count = 0
}
out += string([]byte{c})
last = c
}
count++
}
if count > 0 {
out += fmt.Sprintf(" x %d", count)
}
return out
}
func fmtDataChunks(chunks [][]byte) string {
var out string
for _, chunk := range chunks {
out += fmt.Sprintf("{%q}", fmtDataChunk(chunk))
}
return out
}
func testDataBuffer(t *testing.T, wantBytes []byte, setup func(t *testing.T) *dataBuffer) {
// Run setup, then read the remaining bytes from the dataBuffer and check
// that they match wantBytes. We use different read sizes to check corner
// cases in Read.
for _, readSize := range []int{1, 2, 1 * 1024, 32 * 1024} {
t.Run(fmt.Sprintf("ReadSize=%d", readSize), func(t *testing.T) {
b := setup(t)
buf := make([]byte, readSize)
var gotRead bytes.Buffer
for {
n, err := b.Read(buf)
gotRead.Write(buf[:n])
if err == errReadEmpty {
break
}
if err != nil {
t.Fatalf("error after %v bytes: %v", gotRead.Len(), err)
}
}
if got, want := gotRead.Bytes(), wantBytes; !bytes.Equal(got, want) {
t.Errorf("FinalRead=%q, want %q", fmtDataChunk(got), fmtDataChunk(want))
}
})
}
}
func TestDataBufferAllocation(t *testing.T) {
writes := [][]byte{
bytes.Repeat([]byte("a"), 1*1024-1),
[]byte{'a'},
bytes.Repeat([]byte("b"), 4*1024-1),
[]byte{'b'},
bytes.Repeat([]byte("c"), 8*1024-1),
[]byte{'c'},
bytes.Repeat([]byte("d"), 16*1024-1),
[]byte{'d'},
bytes.Repeat([]byte("e"), 32*1024),
}
var wantRead bytes.Buffer
for _, p := range writes {
wantRead.Write(p)
}
testDataBuffer(t, wantRead.Bytes(), func(t *testing.T) *dataBuffer {
b := &dataBuffer{}
for _, p := range writes {
if n, err := b.Write(p); n != len(p) || err != nil {
t.Fatalf("Write(%q x %d)=%v,%v want %v,nil", p[:1], len(p), n, err, len(p))
}
}
want := [][]byte{
bytes.Repeat([]byte("a"), 1*1024),
bytes.Repeat([]byte("b"), 4*1024),
bytes.Repeat([]byte("c"), 8*1024),
bytes.Repeat([]byte("d"), 16*1024),
bytes.Repeat([]byte("e"), 16*1024),
bytes.Repeat([]byte("e"), 16*1024),
}
if !reflect.DeepEqual(b.chunks, want) {
t.Errorf("dataBuffer.chunks\ngot: %s\nwant: %s", fmtDataChunks(b.chunks), fmtDataChunks(want))
}
return b
})
}
func TestDataBufferAllocationWithExpected(t *testing.T) {
writes := [][]byte{
bytes.Repeat([]byte("a"), 1*1024), // allocates 16KB
bytes.Repeat([]byte("b"), 14*1024),
bytes.Repeat([]byte("c"), 15*1024), // allocates 16KB more
bytes.Repeat([]byte("d"), 2*1024),
bytes.Repeat([]byte("e"), 1*1024), // overflows 32KB expectation, allocates just 1KB
}
var wantRead bytes.Buffer
for _, p := range writes {
wantRead.Write(p)
}
testDataBuffer(t, wantRead.Bytes(), func(t *testing.T) *dataBuffer {
b := &dataBuffer{expected: 32 * 1024}
for _, p := range writes {
if n, err := b.Write(p); n != len(p) || err != nil {
t.Fatalf("Write(%q x %d)=%v,%v want %v,nil", p[:1], len(p), n, err, len(p))
}
}
want := [][]byte{
append(bytes.Repeat([]byte("a"), 1*1024), append(bytes.Repeat([]byte("b"), 14*1024), bytes.Repeat([]byte("c"), 1*1024)...)...),
append(bytes.Repeat([]byte("c"), 14*1024), bytes.Repeat([]byte("d"), 2*1024)...),
bytes.Repeat([]byte("e"), 1*1024),
}
if !reflect.DeepEqual(b.chunks, want) {
t.Errorf("dataBuffer.chunks\ngot: %s\nwant: %s", fmtDataChunks(b.chunks), fmtDataChunks(want))
}
return b
})
}
func TestDataBufferWriteAfterPartialRead(t *testing.T) {
testDataBuffer(t, []byte("cdxyz"), func(t *testing.T) *dataBuffer {
b := &dataBuffer{}
if n, err := b.Write([]byte("abcd")); n != 4 || err != nil {
t.Fatalf("Write(\"abcd\")=%v,%v want 4,nil", n, err)
}
p := make([]byte, 2)
if n, err := b.Read(p); n != 2 || err != nil || !bytes.Equal(p, []byte("ab")) {
t.Fatalf("Read()=%q,%v,%v want \"ab\",2,nil", p, n, err)
}
if n, err := b.Write([]byte("xyz")); n != 3 || err != nil {
t.Fatalf("Write(\"xyz\")=%v,%v want 3,nil", n, err)
}
return b
})
}
// Copyright 2014 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package http2
import (
"errors"
)
// fixedBuffer is an io.ReadWriter backed by a fixed size buffer.
// It never allocates, but moves old data as new data is written.
type fixedBuffer struct {
buf []byte
r, w int
}
var (
errReadEmpty = errors.New("read from empty fixedBuffer")
errWriteFull = errors.New("write on full fixedBuffer")
)
// Read copies bytes from the buffer into p.
// It is an error to read when no data is available.
func (b *fixedBuffer) Read(p []byte) (n int, err error) {
if b.r == b.w {
return 0, errReadEmpty
}
n = copy(p, b.buf[b.r:b.w])
b.r += n
if b.r == b.w {
b.r = 0
b.w = 0
}
return n, nil
}
// Len returns the number of bytes of the unread portion of the buffer.
func (b *fixedBuffer) Len() int {
return b.w - b.r
}
// Write copies bytes from p into the buffer.
// It is an error to write more data than the buffer can hold.
func (b *fixedBuffer) Write(p []byte) (n int, err error) {
// Slide existing data to beginning.
if b.r > 0 && len(p) > len(b.buf)-b.w {
copy(b.buf, b.buf[b.r:b.w])
b.w -= b.r
b.r = 0
}
// Write new data.
n = copy(b.buf[b.w:], p)
b.w += n
if n < len(p) {
err = errWriteFull
}
return n, err
}
// Copyright 2014 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package http2
import (
"reflect"
"testing"
)
var bufferReadTests = []struct {
buf fixedBuffer
read, wn int
werr error
wp []byte
wbuf fixedBuffer
}{
{
fixedBuffer{[]byte{'a', 0}, 0, 1},
5, 1, nil, []byte{'a'},
fixedBuffer{[]byte{'a', 0}, 0, 0},
},
{
fixedBuffer{[]byte{0, 'a'}, 1, 2},
5, 1, nil, []byte{'a'},
fixedBuffer{[]byte{0, 'a'}, 0, 0},
},
{
fixedBuffer{[]byte{'a', 'b'}, 0, 2},
1, 1, nil, []byte{'a'},
fixedBuffer{[]byte{'a', 'b'}, 1, 2},
},
{
fixedBuffer{[]byte{}, 0, 0},
5, 0, errReadEmpty, []byte{},
fixedBuffer{[]byte{}, 0, 0},
},
}
func TestBufferRead(t *testing.T) {
for i, tt := range bufferReadTests {
read := make([]byte, tt.read)
n, err := tt.buf.Read(read)
if n != tt.wn {
t.Errorf("#%d: wn = %d want %d", i, n, tt.wn)
continue
}
if err != tt.werr {
t.Errorf("#%d: werr = %v want %v", i, err, tt.werr)
continue
}
read = read[:n]
if !reflect.DeepEqual(read, tt.wp) {
t.Errorf("#%d: read = %+v want %+v", i, read, tt.wp)
}
if !reflect.DeepEqual(tt.buf, tt.wbuf) {
t.Errorf("#%d: buf = %+v want %+v", i, tt.buf, tt.wbuf)
}
}
}
var bufferWriteTests = []struct {
buf fixedBuffer
write, wn int
werr error
wbuf fixedBuffer
}{
{
buf: fixedBuffer{
buf: []byte{},
},
wbuf: fixedBuffer{
buf: []byte{},
},
},
{
buf: fixedBuffer{
buf: []byte{1, 'a'},
},
write: 1,
wn: 1,
wbuf: fixedBuffer{
buf: []byte{0, 'a'},
w: 1,
},
},
{
buf: fixedBuffer{
buf: []byte{'a', 1},
r: 1,
w: 1,
},
write: 2,
wn: 2,
wbuf: fixedBuffer{
buf: []byte{0, 0},
w: 2,
},
},
{
buf: fixedBuffer{
buf: []byte{},
},
write: 5,
werr: errWriteFull,
wbuf: fixedBuffer{
buf: []byte{},
},
},
}
func TestBufferWrite(t *testing.T) {
for i, tt := range bufferWriteTests {
n, err := tt.buf.Write(make([]byte, tt.write))
if n != tt.wn {
t.Errorf("#%d: wrote %d bytes; want %d", i, n, tt.wn)
continue
}
if err != tt.werr {
t.Errorf("#%d: error = %v; want %v", i, err, tt.werr)
continue
}
if !reflect.DeepEqual(tt.buf, tt.wbuf) {
t.Errorf("#%d: buf = %+v; want %+v", i, tt.buf, tt.wbuf)
}
}
}
...@@ -466,7 +466,6 @@ type stream struct { ...@@ -466,7 +466,6 @@ type stream struct {
resetQueued bool // RST_STREAM queued for write; set by sc.resetStream resetQueued bool // RST_STREAM queued for write; set by sc.resetStream
gotTrailerHeader bool // HEADER frame for trailers was seen gotTrailerHeader bool // HEADER frame for trailers was seen
wroteHeaders bool // whether we wrote headers (not status 100) wroteHeaders bool // whether we wrote headers (not status 100)
reqBuf []byte // if non-nil, body pipe buffer to return later at EOF
trailer http.Header // accumulated trailers trailer http.Header // accumulated trailers
reqTrailer http.Header // handler's Request.Trailer reqTrailer http.Header // handler's Request.Trailer
...@@ -1785,16 +1784,14 @@ func (sc *serverConn) newWriterAndRequest(st *stream, f *MetaHeadersFrame) (*res ...@@ -1785,16 +1784,14 @@ func (sc *serverConn) newWriterAndRequest(st *stream, f *MetaHeadersFrame) (*res
return nil, nil, err return nil, nil, err
} }
if bodyOpen { if bodyOpen {
st.reqBuf = getRequestBodyBuf()
req.Body.(*requestBody).pipe = &pipe{
b: &fixedBuffer{buf: st.reqBuf},
}
if vv, ok := rp.header["Content-Length"]; ok { if vv, ok := rp.header["Content-Length"]; ok {
req.ContentLength, _ = strconv.ParseInt(vv[0], 10, 64) req.ContentLength, _ = strconv.ParseInt(vv[0], 10, 64)
} else { } else {
req.ContentLength = -1 req.ContentLength = -1
} }
req.Body.(*requestBody).pipe = &pipe{
b: &dataBuffer{expected: req.ContentLength},
}
} }
return rw, req, nil return rw, req, nil
} }
...@@ -1890,24 +1887,6 @@ func (sc *serverConn) newWriterAndRequestNoBody(st *stream, rp requestParam) (*r ...@@ -1890,24 +1887,6 @@ func (sc *serverConn) newWriterAndRequestNoBody(st *stream, rp requestParam) (*r
return rw, req, nil return rw, req, nil
} }
var reqBodyCache = make(chan []byte, 8)
func getRequestBodyBuf() []byte {
select {
case b := <-reqBodyCache:
return b
default:
return make([]byte, initialWindowSize)
}
}
func putRequestBodyBuf(b []byte) {
select {
case reqBodyCache <- b:
default:
}
}
// Run on its own goroutine. // Run on its own goroutine.
func (sc *serverConn) runHandler(rw *responseWriter, req *http.Request, handler func(http.ResponseWriter, *http.Request)) { func (sc *serverConn) runHandler(rw *responseWriter, req *http.Request, handler func(http.ResponseWriter, *http.Request)) {
didPanic := true didPanic := true
...@@ -2003,12 +1982,6 @@ func (sc *serverConn) noteBodyReadFromHandler(st *stream, n int, err error) { ...@@ -2003,12 +1982,6 @@ func (sc *serverConn) noteBodyReadFromHandler(st *stream, n int, err error) {
case <-sc.doneServing: case <-sc.doneServing:
} }
} }
if err == io.EOF {
if buf := st.reqBuf; buf != nil {
st.reqBuf = nil // shouldn't matter; field unused by other
putRequestBodyBuf(buf)
}
}
} }
func (sc *serverConn) noteBodyRead(st *stream, n int) { func (sc *serverConn) noteBodyRead(st *stream, n int) {
......
...@@ -1528,8 +1528,7 @@ func (rl *clientConnReadLoop) handleResponse(cs *clientStream, f *MetaHeadersFra ...@@ -1528,8 +1528,7 @@ func (rl *clientConnReadLoop) handleResponse(cs *clientStream, f *MetaHeadersFra
return res, nil return res, nil
} }
buf := new(bytes.Buffer) // TODO(bradfitz): recycle this garbage cs.bufPipe = pipe{b: &dataBuffer{expected: res.ContentLength}}
cs.bufPipe = pipe{b: buf}
cs.bytesRemain = res.ContentLength cs.bytesRemain = res.ContentLength
res.Body = transportResponseBody{cs} res.Body = transportResponseBody{cs}
go cs.awaitRequestCancel(cs.req) go cs.awaitRequestCancel(cs.req)
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment