io, net, http: sendfile support
Speeds up static fileserver, avoiding kernel/userspace copies.
Numbers: downloading 14 MB AppEngine Go SDK with ab (Apache Bench)
with 5 threads:
Before/after numbers:
CPU:
user 0m3.910s
sys 0m23.650s
->
user 0m0.720s
sys 0m4.890s
Time taken for tests: 8.906 seconds
->
Time taken for tests: 8.545 seconds
Percentage of the requests served within a certain time (ms)
50% 44
66% 45
75% 46
80% 46
90% 48
95% 51
98% 59
99% 71
100 74 (longest request)
->
50% 42
66% 43
75% 43
80% 44
90% 46
95% 57
98% 62
99% 63
100% 64 (longest request)
R=iant, gary.burd, rsc, bradfitz
CC=golang-dev
https://golang.org/cl/4543071
diff --git a/src/pkg/http/server.go b/src/pkg/http/server.go
index eb5a3a3..ffeac03 100644
--- a/src/pkg/http/server.go
+++ b/src/pkg/http/server.go
@@ -119,6 +119,27 @@
closeAfterReply bool
}
+type writerOnly struct {
+ io.Writer
+}
+
+func (r *response) ReadFrom(src io.Reader) (n int64, err os.Error) {
+ // Flush before checking r.chunking, as Flush will call
+ // WriteHeader if it hasn't been called yet, and WriteHeader
+ // is what sets r.chunking.
+ r.Flush()
+ if !r.chunking {
+ if rf, ok := r.conn.rwc.(io.ReaderFrom); ok {
+ n, err = rf.ReadFrom(src)
+ r.written += n
+ return
+ }
+ }
+ // Fall back to default io.Copy implementation.
+ // Use wrapper to hide r.ReadFrom from io.Copy.
+ return io.Copy(writerOnly{r}, src)
+}
+
// Create new connection from rwc.
func newConn(rwc net.Conn, handler Handler) (c *conn, err os.Error) {
c = new(conn)
diff --git a/src/pkg/io/io.go b/src/pkg/io/io.go
index 0bc73d6..1ad1129 100644
--- a/src/pkg/io/io.go
+++ b/src/pkg/io/io.go
@@ -303,22 +303,26 @@
// LimitReader returns a Reader that reads from r
// but stops with os.EOF after n bytes.
-func LimitReader(r Reader, n int64) Reader { return &limitedReader{r, n} }
+// The underlying implementation is a *LimitedReader.
+func LimitReader(r Reader, n int64) Reader { return &LimitedReader{r, n} }
-type limitedReader struct {
- r Reader
- n int64
+// A LimitedReader reads from R but limits the amount of
+// data returned to just N bytes. Each call to Read
+// updates N to reflect the new amount remaining.
+type LimitedReader struct {
+ R Reader // underlying reader
+ N int64 // max bytes remaining
}
-func (l *limitedReader) Read(p []byte) (n int, err os.Error) {
- if l.n <= 0 {
+func (l *LimitedReader) Read(p []byte) (n int, err os.Error) {
+ if l.N <= 0 {
return 0, os.EOF
}
- if int64(len(p)) > l.n {
- p = p[0:l.n]
+ if int64(len(p)) > l.N {
+ p = p[0:l.N]
}
- n, err = l.r.Read(p)
- l.n -= int64(n)
+ n, err = l.R.Read(p)
+ l.N -= int64(n)
return
}
diff --git a/src/pkg/net/Makefile b/src/pkg/net/Makefile
index 376e9c6..d4adbff 100644
--- a/src/pkg/net/Makefile
+++ b/src/pkg/net/Makefile
@@ -23,12 +23,13 @@
unixsock.go\
GOFILES_freebsd=\
- newpollserver.go\
+ dnsclient.go\
+ dnsconfig.go\
fd.go\
file.go\
- dnsconfig.go\
- dnsclient.go\
+ newpollserver.go\
port.go\
+ sendfile_stub.go\
sock_bsd.go\
CGOFILES_freebsd=\
@@ -36,27 +37,32 @@
cgo_unix.go\
GOFILES_darwin=\
- newpollserver.go\
+ dnsclient.go\
+ dnsconfig.go\
fd.go\
file.go\
- dnsconfig.go\
- dnsclient.go\
+ newpollserver.go\
port.go\
+ sendfile_stub.go\
sock_bsd.go\
CGOFILES_darwin=\
cgo_bsd.go\
cgo_unix.go\
-
+
GOFILES_linux=\
- newpollserver.go\
+ dnsclient.go\
+ dnsconfig.go\
fd.go\
file.go\
- dnsconfig.go\
- dnsclient.go\
+ newpollserver.go\
port.go\
+ sendfile_linux.go\
sock_linux.go\
+GOFILES_plan9=\
+ sendfile_stub.go\
+
ifeq ($(GOARCH),arm)
# ARM has no cgo, so use the stubs.
GOFILES_linux+=cgo_stub.go
@@ -68,8 +74,9 @@
GOFILES_windows=\
cgo_stub.go\
- resolv_windows.go\
file_windows.go\
+ resolv_windows.go\
+ sendfile_stub.go\
sock_windows.go\
GOFILES+=$(GOFILES_$(GOOS))
diff --git a/src/pkg/net/sendfile_linux.go b/src/pkg/net/sendfile_linux.go
new file mode 100644
index 0000000..6a5a06c
--- /dev/null
+++ b/src/pkg/net/sendfile_linux.go
@@ -0,0 +1,84 @@
+// Copyright 2011 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 net
+
+import (
+ "io"
+ "os"
+ "syscall"
+)
+
+// maxSendfileSize is the largest chunk size we ask the kernel to copy
+// at a time.
+const maxSendfileSize int = 4 << 20
+
+// sendFile copies the contents of r to c using the sendfile
+// system call to minimize copies.
+//
+// if handled == true, sendFile returns the number of bytes copied and any
+// non-EOF error.
+//
+// if handled == false, sendFile performed no work.
+func sendFile(c *netFD, r io.Reader) (written int64, err os.Error, handled bool) {
+ var remain int64 = 1 << 62 // by default, copy until EOF
+
+ lr, ok := r.(*io.LimitedReader)
+ if ok {
+ remain, r = lr.N, lr.R
+ if remain <= 0 {
+ return 0, nil, true
+ }
+ }
+ f, ok := r.(*os.File)
+ if !ok {
+ return 0, nil, false
+ }
+
+ c.wio.Lock()
+ defer c.wio.Unlock()
+ c.incref()
+ defer c.decref()
+ if c.wdeadline_delta > 0 {
+ // This is a little odd that we're setting the timeout
+ // for the entire file but Write has the same issue
+ // (if one slurps the whole file into memory and
+ // do one large Write). At least they're consistent.
+ c.wdeadline = pollserver.Now() + c.wdeadline_delta
+ } else {
+ c.wdeadline = 0
+ }
+
+ dst := c.sysfd
+ src := f.Fd()
+ for remain > 0 {
+ n := maxSendfileSize
+ if int64(n) > remain {
+ n = int(remain)
+ }
+ n, errno := syscall.Sendfile(dst, src, nil, n)
+ if n > 0 {
+ written += int64(n)
+ remain -= int64(n)
+ }
+ if n == 0 && errno == 0 {
+ break
+ }
+ if errno == syscall.EAGAIN && c.wdeadline >= 0 {
+ pollserver.WaitWrite(c)
+ continue
+ }
+ if errno != 0 {
+ // This includes syscall.ENOSYS (no kernel
+ // support) and syscall.EINVAL (fd types which
+ // don't implement sendfile together)
+ err = &OpError{"sendfile", c.net, c.raddr, os.Errno(errno)}
+ break
+ }
+ }
+ if lr != nil {
+ lr.N = remain
+ }
+ return written, err, written > 0
+}
diff --git a/src/pkg/net/sendfile_stub.go b/src/pkg/net/sendfile_stub.go
new file mode 100644
index 0000000..43e8104
--- /dev/null
+++ b/src/pkg/net/sendfile_stub.go
@@ -0,0 +1,14 @@
+// Copyright 2011 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 net
+
+import (
+ "io"
+ "os"
+)
+
+func sendFile(c *netFD, r io.Reader) (n int64, err os.Error, handled bool) {
+ return 0, nil, false
+}
diff --git a/src/pkg/net/sock.go b/src/pkg/net/sock.go
index 5c47e4f..eae7f37 100644
--- a/src/pkg/net/sock.go
+++ b/src/pkg/net/sock.go
@@ -7,6 +7,7 @@
package net
import (
+ "io"
"os"
"reflect"
"syscall"
@@ -153,3 +154,14 @@
func (e *UnknownSocketError) String() string {
return "unknown socket address type " + reflect.TypeOf(e.sa).String()
}
+
+type writerOnly struct {
+ io.Writer
+}
+
+// Fallback implementation of io.ReaderFrom's ReadFrom, when sendfile isn't
+// applicable.
+func genericReadFrom(w io.Writer, r io.Reader) (n int64, err os.Error) {
+ // Use wrapper to hide existing r.ReadFrom from io.Copy.
+ return io.Copy(writerOnly{w}, r)
+}
diff --git a/src/pkg/net/tcpsock.go b/src/pkg/net/tcpsock.go
index 8aeed48..9ee6c14 100644
--- a/src/pkg/net/tcpsock.go
+++ b/src/pkg/net/tcpsock.go
@@ -7,6 +7,7 @@
package net
import (
+ "io"
"os"
"syscall"
)
@@ -95,6 +96,14 @@
return c.fd.Read(b)
}
+// ReadFrom implements the io.ReaderFrom ReadFrom method.
+func (c *TCPConn) ReadFrom(r io.Reader) (int64, os.Error) {
+ if n, err, handled := sendFile(c.fd, r); handled {
+ return n, err
+ }
+ return genericReadFrom(c, r)
+}
+
// Write implements the net.Conn Write method.
func (c *TCPConn) Write(b []byte) (n int, err os.Error) {
if !c.ok() {
