Check in gVisor.

PiperOrigin-RevId: 194583126
Change-Id: Ica1d8821a90f74e7e745962d71801c598c652463

4 files changed, 244 insertions, 0 deletions

diff --git a/pkg/tcpip/link/rawfile/BUILD b/pkg/tcpip/link/rawfile/BUILD
new file mode 100644
index 000000000..4c63af0ea
--- /dev/null
+++ b/pkg/tcpip/link/rawfile/BUILD
@@ -0,0 +1,17 @@
+package(licenses = ["notice"])  # BSD
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "rawfile",
+    srcs = [
+        "blockingpoll_amd64.s",
+        "errors.go",
+        "rawfile_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/tcpip/link/rawfile",
+    visibility = [
+        "//visibility:public",
+    ],
+    deps = ["//pkg/tcpip"],
+)
diff --git a/pkg/tcpip/link/rawfile/blockingpoll_amd64.s b/pkg/tcpip/link/rawfile/blockingpoll_amd64.s
new file mode 100644
index 000000000..88206fc87
--- /dev/null
+++ b/pkg/tcpip/link/rawfile/blockingpoll_amd64.s
@@ -0,0 +1,26 @@
+#include "textflag.h"
+
+// blockingPoll makes the poll() syscall while calling the version of
+// entersyscall that relinquishes the P so that other Gs can run. This is meant
+// to be called in cases when the syscall is expected to block.
+//
+// func blockingPoll(fds unsafe.Pointer, nfds int, timeout int64) (n int, err syscall.Errno)
+TEXT ·blockingPoll(SB),NOSPLIT,$0-40
+	CALL	runtime·entersyscallblock(SB)
+	MOVQ	fds+0(FP), DI
+	MOVQ	nfds+8(FP), SI
+	MOVQ	timeout+16(FP), DX
+	MOVQ	$0x7, AX // SYS_POLL
+	SYSCALL
+	CMPQ	AX, $0xfffffffffffff001
+	JLS	ok
+	MOVQ	$-1, n+24(FP)
+	NEGQ	AX
+	MOVQ	AX, err+32(FP)
+	CALL	runtime·exitsyscall(SB)
+	RET
+ok:
+	MOVQ	AX, n+24(FP)
+	MOVQ	$0, err+32(FP)
+	CALL	runtime·exitsyscall(SB)
+	RET
diff --git a/pkg/tcpip/link/rawfile/errors.go b/pkg/tcpip/link/rawfile/errors.go
new file mode 100644
index 000000000..b6e7b3d71
--- /dev/null
+++ b/pkg/tcpip/link/rawfile/errors.go
@@ -0,0 +1,40 @@
+package rawfile
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+)
+
+var translations = map[syscall.Errno]*tcpip.Error{
+	syscall.EEXIST:        tcpip.ErrDuplicateAddress,
+	syscall.ENETUNREACH:   tcpip.ErrNoRoute,
+	syscall.EINVAL:        tcpip.ErrInvalidEndpointState,
+	syscall.EALREADY:      tcpip.ErrAlreadyConnecting,
+	syscall.EISCONN:       tcpip.ErrAlreadyConnected,
+	syscall.EADDRINUSE:    tcpip.ErrPortInUse,
+	syscall.EADDRNOTAVAIL: tcpip.ErrBadLocalAddress,
+	syscall.EPIPE:         tcpip.ErrClosedForSend,
+	syscall.EWOULDBLOCK:   tcpip.ErrWouldBlock,
+	syscall.ECONNREFUSED:  tcpip.ErrConnectionRefused,
+	syscall.ETIMEDOUT:     tcpip.ErrTimeout,
+	syscall.EINPROGRESS:   tcpip.ErrConnectStarted,
+	syscall.EDESTADDRREQ:  tcpip.ErrDestinationRequired,
+	syscall.ENOTSUP:       tcpip.ErrNotSupported,
+	syscall.ENOTTY:        tcpip.ErrQueueSizeNotSupported,
+	syscall.ENOTCONN:      tcpip.ErrNotConnected,
+	syscall.ECONNRESET:    tcpip.ErrConnectionReset,
+	syscall.ECONNABORTED:  tcpip.ErrConnectionAborted,
+}
+
+// TranslateErrno translate an errno from the syscall package into a
+// *tcpip.Error.
+//
+// Not all errnos are supported and this function will panic on unreconized
+// errnos.
+func TranslateErrno(e syscall.Errno) *tcpip.Error {
+	if err, ok := translations[e]; ok {
+		return err
+	}
+	return tcpip.ErrInvalidEndpointState
+}
diff --git a/pkg/tcpip/link/rawfile/rawfile_unsafe.go b/pkg/tcpip/link/rawfile/rawfile_unsafe.go
new file mode 100644
index 000000000..d3660e1b4
--- /dev/null
+++ b/pkg/tcpip/link/rawfile/rawfile_unsafe.go
@@ -0,0 +1,161 @@
+// Copyright 2016 The Netstack 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 rawfile contains utilities for using the netstack with raw host
+// files on Linux hosts.
+package rawfile
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+)
+
+//go:noescape
+func blockingPoll(fds unsafe.Pointer, nfds int, timeout int64) (n int, err syscall.Errno)
+
+// GetMTU determines the MTU of a network interface device.
+func GetMTU(name string) (uint32, error) {
+	fd, err := syscall.Socket(syscall.AF_UNIX, syscall.SOCK_DGRAM, 0)
+	if err != nil {
+		return 0, err
+	}
+
+	defer syscall.Close(fd)
+
+	var ifreq struct {
+		name [16]byte
+		mtu  int32
+		_    [20]byte
+	}
+
+	copy(ifreq.name[:], name)
+	_, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), syscall.SIOCGIFMTU, uintptr(unsafe.Pointer(&ifreq)))
+	if errno != 0 {
+		return 0, errno
+	}
+
+	return uint32(ifreq.mtu), nil
+}
+
+// NonBlockingWrite writes the given buffer to a file descriptor. It fails if
+// partial data is written.
+func NonBlockingWrite(fd int, buf []byte) *tcpip.Error {
+	var ptr unsafe.Pointer
+	if len(buf) > 0 {
+		ptr = unsafe.Pointer(&buf[0])
+	}
+
+	_, _, e := syscall.RawSyscall(syscall.SYS_WRITE, uintptr(fd), uintptr(ptr), uintptr(len(buf)))
+	if e != 0 {
+		return TranslateErrno(e)
+	}
+
+	return nil
+}
+
+// NonBlockingWrite2 writes up to two byte slices to a file descriptor in a
+// single syscall. It fails if partial data is written.
+func NonBlockingWrite2(fd int, b1, b2 []byte) *tcpip.Error {
+	// If the is no second buffer, issue a regular write.
+	if len(b2) == 0 {
+		return NonBlockingWrite(fd, b1)
+	}
+
+	// We have two buffers. Build the iovec that represents them and issue
+	// a writev syscall.
+	iovec := [...]syscall.Iovec{
+		{
+			Base: &b1[0],
+			Len:  uint64(len(b1)),
+		},
+		{
+			Base: &b2[0],
+			Len:  uint64(len(b2)),
+		},
+	}
+
+	_, _, e := syscall.RawSyscall(syscall.SYS_WRITEV, uintptr(fd), uintptr(unsafe.Pointer(&iovec[0])), uintptr(len(iovec)))
+	if e != 0 {
+		return TranslateErrno(e)
+	}
+
+	return nil
+}
+
+// NonBlockingWriteN writes up to N byte slices to a file descriptor in a
+// single syscall. It fails if partial data is written.
+func NonBlockingWriteN(fd int, bs ...[]byte) *tcpip.Error {
+	iovec := make([]syscall.Iovec, 0, len(bs))
+
+	for _, b := range bs {
+		if len(b) == 0 {
+			continue
+		}
+		iovec = append(iovec, syscall.Iovec{
+			Base: &b[0],
+			Len:  uint64(len(b)),
+		})
+	}
+
+	_, _, e := syscall.RawSyscall(syscall.SYS_WRITEV, uintptr(fd), uintptr(unsafe.Pointer(&iovec[0])), uintptr(len(iovec)))
+	if e != 0 {
+		return TranslateErrno(e)
+	}
+
+	return nil
+}
+
+// BlockingRead reads from a file descriptor that is set up as non-blocking. If
+// no data is available, it will block in a poll() syscall until the file
+// descirptor becomes readable.
+func BlockingRead(fd int, b []byte) (int, *tcpip.Error) {
+	for {
+		n, _, e := syscall.RawSyscall(syscall.SYS_READ, uintptr(fd), uintptr(unsafe.Pointer(&b[0])), uintptr(len(b)))
+		if e == 0 {
+			return int(n), nil
+		}
+
+		event := struct {
+			fd      int32
+			events  int16
+			revents int16
+		}{
+			fd:     int32(fd),
+			events: 1, // POLLIN
+		}
+
+		_, e = blockingPoll(unsafe.Pointer(&event), 1, -1)
+		if e != 0 && e != syscall.EINTR {
+			return 0, TranslateErrno(e)
+		}
+	}
+}
+
+// BlockingReadv reads from a file descriptor that is set up as non-blocking and
+// stores the data in a list of iovecs buffers. If no data is available, it will
+// block in a poll() syscall until the file descirptor becomes readable.
+func BlockingReadv(fd int, iovecs []syscall.Iovec) (int, *tcpip.Error) {
+	for {
+		n, _, e := syscall.RawSyscall(syscall.SYS_READV, uintptr(fd), uintptr(unsafe.Pointer(&iovecs[0])), uintptr(len(iovecs)))
+		if e == 0 {
+			return int(n), nil
+		}
+
+		event := struct {
+			fd      int32
+			events  int16
+			revents int16
+		}{
+			fd:     int32(fd),
+			events: 1, // POLLIN
+		}
+
+		_, e = blockingPoll(unsafe.Pointer(&event), 1, -1)
+		if e != 0 && e != syscall.EINTR {
+			return 0, TranslateErrno(e)
+		}
+	}
+}