503 lines
14 KiB
Go
503 lines
14 KiB
Go
//go:build linux
|
||
|
||
package transport
|
||
|
||
import (
|
||
"encoding/binary"
|
||
"errors"
|
||
"fmt"
|
||
"io"
|
||
"strings"
|
||
"syscall"
|
||
"time"
|
||
|
||
"omnisocketgo/cmd/internal/latencylog"
|
||
"omnisocketgo/cmd/internal/protocol"
|
||
)
|
||
|
||
const (
|
||
linuxTimestampControlBufferSize = 256 // 控制消息缓冲区。
|
||
linuxTXTimestampWaitTimeout = 250 * time.Millisecond // 等待 TX 时间戳的上限。
|
||
linuxTXTimestampPollInterval = time.Millisecond // 轮询 errqueue 的间隔。
|
||
|
||
linuxSOTimestampingNew = 0x41
|
||
linuxSCMTimestampingNew = linuxSOTimestampingNew
|
||
linuxSOEEOriginTimestamping = 4 // timestamping errqueue 事件。
|
||
linuxSCMTstampSnd = 0 // 对应 A_TX_SOFTWARE。
|
||
linuxSCMTstampSched = 1 // 对应 A_TX_SCHED。
|
||
|
||
linuxSOFTimestampingTXSoftware = 1 << 1 // 打开 TX software timestamp。
|
||
linuxSOFTimestampingRXSoftware = 1 << 3 // 打开 RX software timestamp。
|
||
linuxSOFTimestampingSoftware = 1 << 4 // software timestamp 总开关。
|
||
linuxSOFTimestampingOptID = 1 << 7 // 给时间戳关联 ID。
|
||
linuxSOFTimestampingTXSched = 1 << 8 // 打开 TX sched timestamp。
|
||
linuxSOFTimestampingOptTSONLY = 1 << 11 // 只回时间戳。
|
||
linuxSOFTimestampingOptIDTCP = 1 << 16 // 让 TCP 也带 timestamp ID。
|
||
)
|
||
|
||
// 拿到底层 fd，并打开 Linux timestamping。
|
||
func (c *TCPConn) initLinuxTimestamping() error {
|
||
sysConn, ok := c.conn.(interface {
|
||
SyscallConn() (syscall.RawConn, error)
|
||
})
|
||
if !ok {
|
||
return fmt.Errorf("transport: connection does not support SyscallConn")
|
||
}
|
||
|
||
rawConn, err := sysConn.SyscallConn()
|
||
if err != nil || rawConn == nil {
|
||
if err != nil {
|
||
return fmt.Errorf("transport: get syscall conn: %w", err)
|
||
}
|
||
return fmt.Errorf("transport: missing syscall conn")
|
||
}
|
||
|
||
//socket是否可以成功打开 timestamping 取决于内核版本和配置，尝试多个 flag 组合直到成功或遇到非 EINVAL 错误。
|
||
if err := enableLinuxTimestamping(rawConn); err != nil {
|
||
return fmt.Errorf("transport: enable linux timestamping: %w", err)
|
||
}
|
||
//成功打开 timestamping 后，rawConn 就可以用来收 TX/RX 时间戳了。
|
||
c.raw = rawConn
|
||
return nil
|
||
}
|
||
|
||
// 给 socket开权限打开TX software timestamping。
|
||
func enableLinuxTimestamping(rawConn syscall.RawConn) error {
|
||
flagCandidates := []int{ //不同linux版本可能支持不同的 flag 组合，尝试多个组合直到成功。
|
||
linuxSOFTimestampingTXSched |
|
||
linuxSOFTimestampingTXSoftware |
|
||
linuxSOFTimestampingRXSoftware |
|
||
linuxSOFTimestampingSoftware |
|
||
linuxSOFTimestampingOptID | //TCP 协议栈给每个时间戳生成一个序列号
|
||
linuxSOFTimestampingOptIDTCP |
|
||
linuxSOFTimestampingOptTSONLY,
|
||
linuxSOFTimestampingTXSched |
|
||
linuxSOFTimestampingTXSoftware |
|
||
linuxSOFTimestampingRXSoftware |
|
||
linuxSOFTimestampingSoftware |
|
||
linuxSOFTimestampingOptID |
|
||
linuxSOFTimestampingOptTSONLY,
|
||
linuxSOFTimestampingTXSched |
|
||
linuxSOFTimestampingTXSoftware |
|
||
linuxSOFTimestampingRXSoftware |
|
||
linuxSOFTimestampingSoftware |
|
||
linuxSOFTimestampingOptTSONLY,
|
||
}
|
||
|
||
var lastErr error
|
||
for _, flags := range flagCandidates { //尝试不同的 flag 组合，直到成功或遇到非 EINVAL 错误。
|
||
// 内核根据 fd 找到对应的内存结构体（Socket 缓冲区）
|
||
err := rawConn.Control(func(fd uintptr) { //Control 方法保证在回调里 fd 是有效的，可以安全地调用 syscall.SetsockoptInt。
|
||
lastErr = syscall.SetsockoptInt(int(fd), syscall.SOL_SOCKET, linuxSOTimestampingNew, flags)
|
||
})
|
||
if err != nil {
|
||
return err
|
||
}
|
||
if lastErr == nil {
|
||
return nil
|
||
}
|
||
if !errors.Is(lastErr, syscall.EINVAL) {
|
||
return lastErr
|
||
}
|
||
}
|
||
|
||
return lastErr
|
||
}
|
||
|
||
// sendMessageLinux 编码消息、写完整帧，再记录 TX 时间戳。
|
||
func (c *TCPConn) sendMessageLinux(msg protocol.Message) error {
|
||
payload, err := protocol.EncodeMessage(msg)
|
||
if err != nil {
|
||
return fmt.Errorf("protocol: encode message: %w", err)
|
||
}
|
||
|
||
//编码后的消息 payload 前面加 4 字节长度，构成完整帧。
|
||
frame := make([]byte, 4+len(payload))
|
||
binary.BigEndian.PutUint32(frame[:4], uint32(len(payload)))
|
||
copy(frame[4:], payload)
|
||
|
||
if err := c.writeFrameLinux(frame); err != nil {
|
||
return fmt.Errorf("protocol: write frame: %w", err)
|
||
}
|
||
//记录发送延时日志
|
||
c.logTXTimestampEvents(msg)
|
||
return nil
|
||
}
|
||
|
||
// writeFrameLinux 用 sendmsg 写完整帧。
|
||
func (c *TCPConn) writeFrameLinux(frame []byte) error {
|
||
written := 0
|
||
var opErr error
|
||
|
||
err := c.raw.Write(func(fd uintptr) bool {
|
||
if written >= len(frame) {
|
||
return true
|
||
}
|
||
|
||
n, sendErr := syscall.SendmsgN(int(fd), frame[written:], nil, nil, 0)
|
||
switch {
|
||
case sendErr == nil:
|
||
if n <= 0 {
|
||
opErr = io.ErrShortWrite
|
||
return true
|
||
}
|
||
written += n
|
||
return written >= len(frame)
|
||
case errors.Is(sendErr, syscall.EAGAIN), errors.Is(sendErr, syscall.EWOULDBLOCK):
|
||
return false
|
||
default:
|
||
opErr = sendErr
|
||
return true
|
||
}
|
||
})
|
||
if err != nil {
|
||
if isRawConnNotPollable(err) {
|
||
return writeFullFallback(c.conn, frame[written:])
|
||
}
|
||
return err
|
||
}
|
||
if opErr != nil {
|
||
if isRawConnNotPollable(opErr) {
|
||
return writeFullFallback(c.conn, frame[written:])
|
||
}
|
||
return opErr
|
||
}
|
||
if written != len(frame) {
|
||
return io.ErrShortWrite
|
||
}
|
||
|
||
return nil
|
||
}
|
||
|
||
// 把 A_TX_SCHED / A_TX_SOFTWARE 写入日志。(发送过程中）
|
||
func (c *TCPConn) logTXTimestampEvents(msg protocol.Message) {
|
||
timestamps := c.collectTXTimestampEvents()
|
||
|
||
if ts, ok := timestamps[latencylog.EventATXSched]; ok {
|
||
latencylog.LogMessageEventAt(c.logger, c.nodeRole, c.nodeID, latencylog.EventATXSched, ts, msg)
|
||
}
|
||
if ts, ok := timestamps[latencylog.EventATXSoftware]; ok {
|
||
latencylog.LogMessageEventAt(c.logger, c.nodeRole, c.nodeID, latencylog.EventATXSoftware, ts, msg)
|
||
}
|
||
}
|
||
|
||
// 在 errqueue 里等两类 TX 时间戳。
|
||
func (c *TCPConn) collectTXTimestampEvents() map[string]int64 {
|
||
timestamps := make(map[string]int64, 2)
|
||
//设置合理等待上限
|
||
deadline := time.Now().Add(linuxTXTimestampWaitTimeout)
|
||
|
||
//轮询 errqueue 直到拿到两类时间戳，或超时，或遇到非 EAGAIN 错误。
|
||
for len(timestamps) < 2 && time.Now().Before(deadline) {
|
||
eventName, ts, err := c.recvTXTimestampOnce()
|
||
if err != nil {
|
||
if isWouldBlock(err) {
|
||
time.Sleep(linuxTXTimestampPollInterval)
|
||
continue
|
||
}
|
||
break
|
||
}
|
||
if eventName == "" || ts <= 0 {
|
||
continue
|
||
}
|
||
if _, exists := timestamps[eventName]; !exists {
|
||
timestamps[eventName] = ts
|
||
}
|
||
}
|
||
|
||
return timestamps
|
||
}
|
||
|
||
// recvTXTimestampOnce 从 errqueue 读一次时间戳事件。
|
||
func (c *TCPConn) recvTXTimestampOnce() (string, int64, error) {
|
||
var (
|
||
eventName string // 事件名，例如 A_TX_SCHED 或 A_TX_SOFTWARE。
|
||
tsUnixNS int64 // 时间戳的 UnixNano 表示。
|
||
opErr error
|
||
)
|
||
|
||
err := c.raw.Control(func(fd uintptr) {
|
||
//设置足够大的 oob buffer 来接收控制消息，调用 recvmsg 从 errqueue 读一条消息。
|
||
oob := make([]byte, linuxTimestampControlBufferSize)
|
||
//recvmsg 的 flags 里必须带 MSG_ERRQUEUE，才能从 errqueue 里读消息，非阻塞模式下如果没有消息可读会返回 EAGAIN。
|
||
_, oobn, _, _, recvErr := syscall.Recvmsg(int(fd), nil, oob, syscall.MSG_ERRQUEUE|syscall.MSG_DONTWAIT)
|
||
if recvErr != nil {
|
||
opErr = recvErr
|
||
return
|
||
}
|
||
//解析控制消息，看看是不是我们关心的 TX 时间戳事件，如果是就拿到事件名和时间戳。
|
||
eventName, tsUnixNS = parseTXTimestampControlMessages(oob[:oobn])
|
||
})
|
||
if err != nil {
|
||
return "", 0, err
|
||
}
|
||
if opErr != nil {
|
||
return "", 0, opErr
|
||
}
|
||
|
||
return eventName, tsUnixNS, nil //如果成功拿到时间戳事件，eventName 会是 A_TX_SCHED 或 A_TX_SOFTWARE 之一，tsUnixNS 是对应的时间戳；如果没有拿到事件或时间戳无效，eventName 会是空字符串，tsUnixNS 会是 0。
|
||
}
|
||
|
||
// 把底层时间戳映射成日志事件名。
|
||
func parseTXTimestampControlMessages(oob []byte) (string, int64) {
|
||
if len(oob) == 0 {
|
||
return "", 0
|
||
}
|
||
//解析控制消息，看看是不是我们关心的 TX 时间戳事件，如果是就拿到事件名和时间戳。
|
||
controlMessages, err := syscall.ParseSocketControlMessage(oob)
|
||
if err != nil {
|
||
return "", 0
|
||
}
|
||
|
||
var (
|
||
tsUnixNS int64 //时间戳的 UnixNano 表示。
|
||
tsKind uint32 //extended err里，告诉我们这个时间戳是 sched 还是 software。
|
||
hasTS bool // 是否拿到时间戳了。
|
||
hasKind bool // 是否拿到时间戳类型了。
|
||
)
|
||
//一个 recvmsg 可能会收到多个控制消息，循环找我们关心的时间戳事件，拿到时间戳和事件类型。
|
||
for _, controlMessage := range controlMessages {
|
||
switch {
|
||
case controlMessage.Header.Level == syscall.SOL_SOCKET && controlMessage.Header.Type == linuxSCMTimestampingNew:
|
||
if ts := parseSCMTimestampingData(controlMessage.Data); ts > 0 {
|
||
tsUnixNS = ts
|
||
hasTS = true
|
||
}
|
||
case isSocketExtendedErr(controlMessage): //判断时间戳是否进入了errqueue，
|
||
if info, ok := parseSocketExtendedErrInfo(controlMessage.Data); ok {
|
||
tsKind = info //时间戳类型被内核放在 extended err 的附加信息里，解析出来。
|
||
hasKind = true
|
||
}
|
||
}
|
||
}
|
||
|
||
if !hasTS || !hasKind {
|
||
return "", 0
|
||
}
|
||
|
||
switch tsKind { //把内核的时间戳类型映射成日志事件名。（记录时只关心 sched 和 software 两类时间戳）
|
||
case linuxSCMTstampSched:
|
||
return latencylog.EventATXSched, tsUnixNS
|
||
case linuxSCMTstampSnd:
|
||
return latencylog.EventATXSoftware, tsUnixNS
|
||
default:
|
||
return "", 0
|
||
}
|
||
}
|
||
|
||
// 判断控制消息是否来自 socket extended err。
|
||
// 内核产生的时间戳并不会混合在普通的数据流里，而是被包装成一种特殊的“错误消息”丢进 Error Queue。
|
||
func isSocketExtendedErr(controlMessage syscall.SocketControlMessage) bool {
|
||
switch {
|
||
case controlMessage.Header.Level == syscall.SOL_IP && controlMessage.Header.Type == syscall.IP_RECVERR:
|
||
return true
|
||
case controlMessage.Header.Level == syscall.SOL_IPV6 && controlMessage.Header.Type == syscall.IPV6_RECVERR:
|
||
return true
|
||
default:
|
||
return false
|
||
}
|
||
}
|
||
|
||
// 从 socket extended err 的数据里取 origin timestamping 信息。
|
||
func parseSocketExtendedErrInfo(data []byte) (uint32, bool) {
|
||
if len(data) < 16 {
|
||
return 0, false
|
||
}
|
||
if data[4] != linuxSOEEOriginTimestamping {
|
||
return 0, false
|
||
}
|
||
|
||
return binary.NativeEndian.Uint32(data[8:12]), true
|
||
}
|
||
|
||
// 读一条完整消息，并记录 B_RX_SOFTWARE。
|
||
func (c *TCPConn) receiveMessageLinux() (protocol.Message, error) {
|
||
payload, rxTimestamp, err := c.readFrameLinux()
|
||
if err != nil {
|
||
return protocol.Message{}, fmt.Errorf("protocol: read frame: %w", err)
|
||
}
|
||
|
||
msg, err := protocol.DecodeMessage(payload)
|
||
if err != nil {
|
||
return protocol.Message{}, fmt.Errorf("protocol: decode message: %w", err)
|
||
}
|
||
|
||
if rxTimestamp > 0 {
|
||
latencylog.LogMessageEventAt(c.logger, c.nodeRole, c.nodeID, latencylog.EventBRXSoftware, rxTimestamp, msg)
|
||
}
|
||
|
||
return msg, nil
|
||
}
|
||
|
||
// readFrameLinux 先读 4 字节长度，再读整条 payload。
|
||
func (c *TCPConn) readFrameLinux() ([]byte, int64, error) {
|
||
var frameHeader [4]byte
|
||
rxTimestamp, err := c.readFullLinux(frameHeader[:])
|
||
if err != nil {
|
||
return nil, rxTimestamp, err
|
||
}
|
||
|
||
size := binary.BigEndian.Uint32(frameHeader[:])
|
||
switch {
|
||
case size == 0:
|
||
return nil, rxTimestamp, protocol.ErrInvalidFrameLength
|
||
case size > protocol.MaxFrameSize:
|
||
return nil, rxTimestamp, protocol.ErrFrameTooLarge
|
||
}
|
||
|
||
payload := make([]byte, int(size))
|
||
bodyTimestamp, err := c.readFullLinux(payload)
|
||
if rxTimestamp == 0 {
|
||
rxTimestamp = bodyTimestamp
|
||
}
|
||
if err != nil {
|
||
return nil, rxTimestamp, err
|
||
}
|
||
|
||
return payload, rxTimestamp, nil
|
||
}
|
||
|
||
// 读满 buf，并保留首个 RX_SOFTWARE（返回进入tcp协议栈的时间戳）。
|
||
func (c *TCPConn) readFullLinux(buf []byte) (int64, error) {
|
||
if len(buf) == 0 {
|
||
return 0, nil
|
||
}
|
||
|
||
var (
|
||
offset int
|
||
firstRXTime int64
|
||
)
|
||
|
||
for offset < len(buf) {
|
||
n, rxTimestamp, err := c.recvmsgLinux(buf[offset:])
|
||
if firstRXTime == 0 && rxTimestamp > 0 {
|
||
firstRXTime = rxTimestamp
|
||
}
|
||
if err != nil {
|
||
if isRawConnNotPollable(err) {
|
||
if fallbackErr := readFullFallback(c.conn, buf[offset:]); fallbackErr != nil {
|
||
if errors.Is(fallbackErr, io.EOF) && offset > 0 {
|
||
return firstRXTime, io.ErrUnexpectedEOF
|
||
}
|
||
return firstRXTime, fallbackErr
|
||
}
|
||
return firstRXTime, nil
|
||
}
|
||
if errors.Is(err, io.EOF) && offset > 0 {
|
||
return firstRXTime, io.ErrUnexpectedEOF
|
||
}
|
||
return firstRXTime, err
|
||
}
|
||
|
||
offset += n
|
||
}
|
||
|
||
return firstRXTime, nil
|
||
}
|
||
|
||
// recvmsgLinux 用 recvmsg 同时读取数据和控制消息。
|
||
func (c *TCPConn) recvmsgLinux(buf []byte) (int, int64, error) {
|
||
var (
|
||
n int
|
||
rxTimeNS int64
|
||
opErr error
|
||
)
|
||
|
||
err := c.raw.Read(func(fd uintptr) bool {
|
||
oob := make([]byte, linuxTimestampControlBufferSize)
|
||
readN, oobN, _, _, recvErr := syscall.Recvmsg(int(fd), buf, oob, 0)
|
||
switch {
|
||
case recvErr == nil:
|
||
if readN == 0 {
|
||
opErr = io.EOF
|
||
return true
|
||
}
|
||
n = readN
|
||
rxTimeNS = parseRXTimestampControlMessages(oob[:oobN])
|
||
return true
|
||
case errors.Is(recvErr, syscall.EAGAIN), errors.Is(recvErr, syscall.EWOULDBLOCK):
|
||
return false
|
||
default:
|
||
opErr = recvErr
|
||
return true
|
||
}
|
||
})
|
||
if err != nil {
|
||
return 0, 0, err
|
||
}
|
||
if opErr != nil {
|
||
return 0, 0, opErr
|
||
}
|
||
|
||
return n, rxTimeNS, nil
|
||
}
|
||
|
||
// 从控制消息里取 RX_SOFTWARE。
|
||
func parseRXTimestampControlMessages(oob []byte) int64 {
|
||
if len(oob) == 0 {
|
||
return 0
|
||
}
|
||
|
||
controlMessages, err := syscall.ParseSocketControlMessage(oob)
|
||
if err != nil {
|
||
return 0
|
||
}
|
||
|
||
for _, controlMessage := range controlMessages {
|
||
if controlMessage.Header.Level != syscall.SOL_SOCKET || controlMessage.Header.Type != linuxSCMTimestampingNew {
|
||
continue
|
||
}
|
||
|
||
if ts := parseSCMTimestampingData(controlMessage.Data); ts > 0 {
|
||
return ts
|
||
}
|
||
}
|
||
|
||
return 0
|
||
}
|
||
|
||
// 取第一个非零 timespec。
|
||
func parseSCMTimestampingData(data []byte) int64 {
|
||
const timespec64Size = 16
|
||
|
||
for offset := 0; offset+timespec64Size <= len(data); offset += timespec64Size {
|
||
sec := int64(binary.NativeEndian.Uint64(data[offset : offset+8]))
|
||
nsec := int64(binary.NativeEndian.Uint64(data[offset+8 : offset+16]))
|
||
if sec == 0 && nsec == 0 {
|
||
continue
|
||
}
|
||
return sec*int64(time.Second) + nsec
|
||
}
|
||
|
||
return 0
|
||
}
|
||
|
||
// 判断错误是否是 EAGAIN 或 EWOULDBLOCK。
|
||
func isWouldBlock(err error) bool {
|
||
return errors.Is(err, syscall.EAGAIN) || errors.Is(err, syscall.EWOULDBLOCK)
|
||
}
|
||
|
||
func isRawConnNotPollable(err error) bool {
|
||
return err != nil && strings.Contains(err.Error(), "not pollable")
|
||
}
|
||
|
||
func writeFullFallback(w io.Writer, buf []byte) error {
|
||
for len(buf) > 0 {
|
||
n, err := w.Write(buf)
|
||
if err != nil {
|
||
return err
|
||
}
|
||
if n <= 0 {
|
||
return io.ErrShortWrite
|
||
}
|
||
buf = buf[n:]
|
||
}
|
||
|
||
return nil
|
||
}
|
||
|
||
func readFullFallback(r io.Reader, buf []byte) error {
|
||
_, err := io.ReadFull(r, buf)
|
||
return err
|
||
}
|