1 前言
linuxptp 是linux平台上根据IEEE standard 1588的一个Precision Time Protocol (PTP)协议实现.
更多的特性介绍可以参考页面 linuxptp feature
这里我们尝试从功能和源码的角度进行分析
2 源码分析
2.1 从phc_ctl看ptp时间查询
phc_ctl.c
phc.c
sk.c
utils.c
sequenceDiagram
participant User as user
box phc_ctrl
participant Main as phc_ctl.c<br>main
participant RunCmds as phc_ctl.c<br>run_cmds
participant GetCmd as phc_ctl.c<br>get_command_function
participant DoGet as phc_ctl.c<br>do_get
participant PosixOpen as utils.c<br>posix_clock_open
participant PhcOpen as phc.c<br>phc_open
participant SK as sk.c
end
participant Kernel as Linux kernel
participant PHC as PTP hardware clock
User->>Main: run commands phc_ctl /dev/ptp0 get
Main->>PosixOpen: posix_clock_open("/dev/ptp0", &junk)
alt device is PHC device path
PosixOpen->>PhcOpen: phc_open(device)
PhcOpen->>Kernel: open device file
Kernel-->>PhcOpen: return clockid_t
else device is net interface
PosixOpen->>SK: sk_get_ts_info(device, &ts_info)
SK->>Kernel: ioctl
Kernel-->>PosixOpen: return ethtool_ts_info(X:phc_index)
PosixOpen->>PhcOpen: phc_open("/dev/ptpX")
PhcOpen->>Kernel: open device file
Kernel-->>PhcOpen: return clockid_t
end
PosixOpen-->>Main: return clockid_t
Main->>RunCmds: run_cmds(clkid, cmdc, cmdv)
RunCmds->>GetCmd: get_command_function("get")
GetCmd-->>RunCmds: return do_get func ptr
RunCmds->>DoGet: do_get(clkid, cmdc-i, &cmdv[i])
DoGet->>Kernel: clock_gettime(clkid, &ts)
Kernel->>PHC: read hardware time
PHC-->>Kernel: return time
Kernel-->>DoGet: return time struct
DoGet->>DoGet: serialize time information
DoGet-->>RunCmds: return 0
RunCmds-->>Main: return result
Main-->>User: show time information
关键步骤为
phc_open打开ptp设备, 拿到clkid, clock_gettime获取时间
2.2 E2E mode
-
t1: Master发送SYNC报文的时间点
-
t2: Slave收到SYNC报文的时间点, Master通过FOLLOW_UP的payload把t1发送给Slave
-
t3: Slave发送DELAY_REQ报文的时间点
-
t4: Master收到DELAY_REQ报文的时间点, Master通过DELAY_RESP的payload把t3发送给Slave
sequenceDiagram
participant Master as server(Master)
participant Slave as client(Slave)
Master->>Slave: SYNC
Note over Master,Slave: t1: send timestamps
Note over Master,Slave: t2: receive timestamps
alt 2 step clock
Master->>Slave: FOLLOW_UP (include t1)
end
Slave->>Master: DELAY_REQ
Note over Master,Slave: t3: send timestamps
Note over Master,Slave: t4: receive timestamps
Master->>Slave: DELAY_RESP (include t4)
Note over Slave: Calculate path delays and clock offsets
详细的序列图
sequenceDiagram
participant Master as 主时钟(Master)
participant MPort as 主时钟端口
participant SPort as 从时钟端口
participant Slave as 从时钟(Slave)
%% rect rgb(0xE6, 0xE6, 0xE6)
note right of Master: SYNC 消息流程
Master->>MPort: port_tx_sync()
activate MPort
MPort->>MPort: msg_allocate()
MPort->>MPort: 设置 SYNC 消息头
MPort->>SPort: port_prepare_and_send(TRANS_EVENT)
MPort-->>Master: 返回
deactivate MPort
%% SYNC 消息处理
SPort->>SPort: bc_event() 检测消息
activate SPort
SPort->>SPort: msg_type(msg) == SYNC
SPort->>SPort: process_sync(p, msg)
SPort->>Slave: clock_synchronize()
SPort-->>SPort: 返回
deactivate SPort
%% end
note right of Master: FOLLOW_UP 消息流程
%% FOLLOW_UP 消息流程（2 step clock）
Master->>MPort: port_tx_sync() 继续执行
activate MPort
MPort->>MPort: msg_allocate() 创建 FOLLOW_UP
MPort->>MPort: 设置 FOLLOW_UP 消息头
MPort->>MPort: fup->follow_up.preciseOriginTimestamp = tmv_to_Timestamp(msg->hwts.ts)
MPort->>SPort: port_prepare_and_send(TRANS_GENERAL)
MPort-->>Master: 返回
deactivate MPort
%% FOLLOW_UP 消息处理
SPort->>SPort: bc_event() 检测消息
activate SPort
SPort->>SPort: msg_type(msg) == FOLLOW_UP
SPort->>SPort: process_follow_up(p, msg)
SPort-->>SPort: 返回
deactivate SPort
note left of Slave: DELAY_REQUEST 消息流程
%% DELAY_REQUEST 消息流程
Slave->>SPort: port_delay_request()
activate SPort
SPort->>SPort: msg_allocate()
SPort->>SPort: 设置 DELAY_REQ 消息头
SPort->>MPort: port_prepare_and_send(TRANS_EVENT)
SPort->>SPort: TAILQ_INSERT_HEAD(&p->delay_req, msg, list)
SPort-->>Slave: 返回
deactivate SPort
%% DELAY_REQUEST 消息处理
MPort->>MPort: bc_event() 检测消息
activate MPort
MPort->>MPort: msg_type(msg) == DELAY_REQ
MPort->>MPort: process_delay_req(p, msg)
note right of Master: DELAY_RESPONSE 消息流程
MPort->>MPort: msg_allocate() 创建响应
MPort->>MPort: 设置 DELAY_RESP 消息头
MPort->>MPort: msg->delay_resp.receiveTimestamp = tmv_to_Timestamp(m->hwts.ts)
MPort->>SPort: port_prepare_and_send(TRANS_GENERAL)
MPort-->>MPort: 返回
deactivate MPort
%% DELAY_RESPONSE 消息处理
SPort->>SPort: bc_event() 检测消息
activate SPort
SPort->>SPort: msg_type(msg) == DELAY_RESP
SPort->>SPort: process_delay_resp(p, msg)
SPort->>SPort: 在 p->delay_req 中查找匹配请求
SPort->>SPort: 提取时间戳 (t3, t4)
SPort->>Slave: clock_path_delay(p->clock, t3, t4c)
SPort->>SPort: TAILQ_REMOVE(&p->delay_req, req, list)
SPort->>SPort: msg_put(req)
SPort-->>SPort: 返回
deactivate SPort
以下以UDP port为例
2.2.1 Master Send
Master在port_tx_sync流程中最终通过sk_receive获取到SYNC报文的发送时间点t1
// port.c
int port_tx_sync(struct port *p, struct address *dst, uint16_t sequence_id)
{
// send SYNC
err = port_prepare_and_send(p, msg, event);
// send FollowUp
fup->follow_up.preciseOriginTimestamp = tmv_to_Timestamp(msg->hwts.ts);
err = port_prepare_and_send(p, fup, TRANS_GENERAL);
}
int port_prepare_and_send(struct port *p, struct ptp_message *msg,
enum transport_event event)
{
cnt = transport_send(p->trp, &p->fda, event, msg);
}
// transport.c
int transport_send(struct transport *t, struct fdarray *fda,
enum transport_event event, struct ptp_message *msg)
{
int len = ntohs(msg->header.messageLength);
return t->send(t, fda, event, 0, msg, len, NULL, &msg->hwts);
}
// udp.c
struct transport *udp_transport_create(void)
{
struct udp *udp = calloc(1, sizeof(*udp));
if (!udp)
return NULL;
udp->t.close = udp_close;
udp->t.open = udp_open;
udp->t.recv = udp_recv;
udp->t.send = udp_send; //
udp->t.release = udp_release;
udp->t.physical_addr = udp_physical_addr;
udp->t.protocol_addr = udp_protocol_addr;
return &udp->t;
}
static int udp_send(struct transport *t, struct fdarray *fda,
enum transport_event event, int peer, void *buf, int len,
struct address *addr, struct hw_timestamp *hwts)
{
return event == TRANS_EVENT ? sk_receive(fd, junk, len, NULL, hwts, MSG_ERRQUEUE) : cnt;
}
// sk.c
int sk_receive(int fd, void *buf, int buflen,
struct address *addr, struct hw_timestamp *hwts, int flags)
{
cnt = recvmsg(fd, &msg, flags);
for (cm = CMSG_FIRSTHDR(&msg); cm != NULL; cm = CMSG_NXTHDR(&msg, cm)) {
switch (hwts->type) {
case TS_SOFTWARE:
hwts->ts = timespec_to_tmv(ts[0]);
break;
case TS_HARDWARE:
case TS_ONESTEP:
case TS_P2P1STEP:
hwts->ts = timespec_to_tmv(ts[2]); // hwts->ts
break;
case TS_LEGACY_HW:
hwts->ts = timespec_to_tmv(ts[1]);
break;
}
}
}
2.2.2 Slave Recv
2.2.2.1 Recv
通过poll收取到SYNC,并通过sk_receive获取到SYNC收到的时间点t2
// clock.c
int clock_poll(struct clock *c)
{
cnt = poll(c->pollfd, (c->nports + 2) * N_CLOCK_PFD, -1);
LIST_FOREACH(p, &c->ports, list) {
/* Let the ports handle their events. */
for (i = 0; i < N_POLLFD; i++) {
if (cur[i].revents & (POLLIN|POLLPRI|POLLERR)) {
event = port_event(p, i); //
}
// port.c
static enum fsm_event bc_event(struct port *p, int fd_index)
{
cnt = transport_recv(p->trp, fd, msg);
}
// udp.c
static int udp_recv(struct transport *t, int fd, void *buf, int buflen,
struct address *addr, struct hw_timestamp *hwts)
{
return sk_receive(fd, buf, buflen, addr, hwts, MSG_DONTWAIT); // 记录hwts.ts
}
2.2.2.2 Recv handle
/* receive SYNC */
// port.c
static enum fsm_event bc_event(struct port *p, int fd_index)
case SYNC:
process_sync(p, msg);
// *****receive SYNC*****
void process_sync(struct port *p, struct ptp_message *m)
{
//这里应该是 p->syfu == SF_EMPTY，进而 event = SYNC_MISMATCH，这两个下面的 port_syfufsm() 会用到
if (p->syfu == SF_HAVE_FUP &&
fup_sync_ok(p->last_syncfup, m) &&
p->last_syncfup->header.sequenceId == m->header.sequenceId) {
event = SYNC_MATCH;
} else {
event = SYNC_MISMATCH; //
}
port_syfufsm(p, event, m);
}
// 用来处理 SYNC 和 FOLLOW_UP 消息的乱序问题的状态机
static void port_syfufsm(struct port *p, enum syfu_event event,
struct ptp_message *m)
{
switch (p->syfu) {
case SF_EMPTY:
switch (event) {
case SYNC_MISMATCH:
msg_get(m);
p->last_syncfup = m;
p->syfu = SF_HAVE_SYNC; //
}
// *****receive FOLLOWUP*****
void process_follow_up(struct port *p, struct ptp_message *m)
{
if (p->follow_up_info) {
struct follow_up_info_tlv *fui = follow_up_info_extract(m);
if (!fui)
return;
clock_follow_up_info(p->clock, fui);
}
//接上一个，p->syfu == SF_HAVE_SYNC，进而 event = FUP_MATCH
if (p->syfu == SF_HAVE_SYNC &&
p->last_syncfup->header.sequenceId == m->header.sequenceId) {
event = FUP_MATCH; //
} else {
event = FUP_MISMATCH;
}
port_syfufsm(p, event, m);
}
//
static void port_syfufsm(struct port *p, enum syfu_event event,
struct ptp_message *m)
{
struct ptp_message *syn, *fup;
case SF_HAVE_SYNC:
switch (event) {
case FUP_MATCH:
//获取上一个 msg，在上面一个里面保存了 p->last_syncfup = m;
syn = p->last_syncfup;
// syn->hwts.ts 这个应该是 t2，也就是 sync msg 到达 slave 的时间。m->ts.pdu 应该是 t1，也就是 follow 携带来的 ts。
port_synchronize(p, syn->header.sequenceId,
syn->hwts.ts, m->ts.pdu,
syn->header.correction,
m->header.correction,
m->header.logMessageInterval); //
msg_put(p->last_syncfup);
//这里改了 p->syfu，这样就形成了一个循环
p->syfu = SF_EMPTY;
break;
}
break;
}
}
static void port_synchronize(struct port *p,
uint16_t seqid,
tmv_t ingress_ts,
struct timestamp origin_ts,
Integer64 correction1, Integer64 correction2,
Integer8 sync_interval)
{
enum servo_state state, last_state;
tmv_t t1, t1c, t2, c1, c2;
port_set_sync_rx_tmo(p);
//fup 携带的 t1
t1 = timestamp_to_tmv(origin_ts);
//ingress 获得 t2
t2 = ingress_ts;
c1 = correction_to_tmv(correction1);
c2 = correction_to_tmv(correction2);
t1c = tmv_add(t1, tmv_add(c1, c2));
state = clock_synchronize(p->clock, t2, t1c);
switch (p->state) {
case PS_UNCALIBRATED:
case PS_SLAVE:
//这里主要是 record t1，t2
monitor_sync(p->slave_event_monitor,
clock_parent_identity(p->clock), seqid,
t1, tmv_add(c1, c2), t2);
break;
default:
break;
}
}
2.2.3 slave send DELAY_REQ
int port_delay_request(struct port *p)
{
struct ptp_message *msg;
/* Time to send a new request, forget current pdelay resp and fup */
if (p->peer_delay_resp) {
msg_put(p->peer_delay_resp);
p->peer_delay_resp = NULL;
}
if (p->peer_delay_fup) {
msg_put(p->peer_delay_fup);
p->peer_delay_fup = NULL;
}
if (p->delayMechanism == DM_P2P) {
return port_pdelay_request(p);
}
msg = msg_allocate();
if (!msg) {
return -1;
}
msg->hwts.type = p->timestamping;
msg->header.tsmt = DELAY_REQ | p->transportSpecific;
msg->header.ver = PTP_VERSION;
msg->header.messageLength = sizeof(struct delay_req_msg);
msg->header.domainNumber = clock_domain_number(p->clock);
msg->header.correction = -p->asymmetry;
msg->header.sourcePortIdentity = p->portIdentity;
msg->header.sequenceId = p->seqnum.delayreq++;
msg->header.control = CTL_DELAY_REQ;
msg->header.logMessageInterval = 0x7f;
// 调用 sk_receive 保存t3
if (port_prepare_and_send(p, msg, TRANS_EVENT)) {
pr_err("port %hu: send delay request failed", portnum(p));
goto out;
}
}
2.2.4 master send DELAY_RESP
master收到DELAY_REQ会答复DELAY_RESP
static int process_delay_req(struct port *p, struct ptp_message *m)
{
int err, nsm, saved_seqnum_sync;
struct ptp_message *msg;
nsm = port_nsm_reply(p, m);
if (!nsm && p->state != PS_MASTER && p->state != PS_GRAND_MASTER) {
return 0;
}
if (p->delayMechanism == DM_P2P) {
pr_warning("port %hu: delay request on P2P port", portnum(p));
return 0;
}
msg = msg_allocate();
if (!msg) {
return -1;
}
msg->hwts.type = p->timestamping;
msg->header.tsmt = DELAY_RESP | p->transportSpecific;
msg->header.ver = PTP_VERSION;
msg->header.messageLength = sizeof(struct delay_resp_msg);
msg->header.domainNumber = m->header.domainNumber;
msg->header.correction = m->header.correction;
msg->header.sourcePortIdentity = p->portIdentity;
msg->header.sequenceId = m->header.sequenceId;
msg->header.control = CTL_DELAY_RESP;
msg->header.logMessageInterval = p->logMinDelayReqInterval;
// 将 t4 放到 Delay_Resp msg 中
msg->delay_resp.receiveTimestamp = tmv_to_Timestamp(m->hwts.ts);
msg->delay_resp.requestingPortIdentity = m->header.sourcePortIdentity;
err = port_prepare_and_send(p, msg, TRANS_GENERAL);
if (err) {
pr_err("port %hu: send delay response failed", portnum(p));
goto out;
}
}
2.2.5 slave receive DELAY_RESP
void process_delay_resp(struct port *p, struct ptp_message *m)
{
struct delay_resp_msg *rsp = &m->delay_resp;
struct ptp_message *req;
tmv_t c3, t3, t4, t4c;
// 找到发送 Delay_Req 的那个 msg
TAILQ_FOREACH(req, &p->delay_req, list) {
if (rsp->hdr.sequenceId == ntohs(req->delay_req.hdr.sequenceId)) {
break;
}
}
c3 = correction_to_tmv(m->header.correction);
//得到 t3
t3 = req->hwts.ts;
//t4 从 Delay_Resp msg 中获得
t4 = timestamp_to_tmv(m->ts.pdu);
t4c = tmv_sub(t4, c3);
//record t3 和 t4
monitor_delay(p->slave_event_monitor, clock_parent_identity(p->clock),
m->header.sequenceId, t3, c3, t4);
clock_path_delay(p->clock, t3, t4c); //
TAILQ_REMOVE(&p->delay_req, req, list);
msg_put(req);
if (p->logMinDelayReqInterval == rsp->hdr.logMessageInterval) {
return;
}
if (msg_unicast(m)) {
/* Unicast responses have logMinDelayReqInterval set to 0x7F. */
return;
}
}
// 计算路径延迟
void clock_path_delay(struct clock *c, tmv_t req, tmv_t rx)
{
tsproc_up_ts(c->tsproc, req, rx);
if (tsproc_update_delay(c->tsproc, &c->path_delay))
return;
c->cur.meanPathDelay = tmv_to_TimeInterval(c->path_delay);
if (c->stats.delay)
stats_add_value(c->stats.delay, tmv_dbl(c->path_delay));
}
2.3 common logic
2.3.1 adjust time
在clock_synchronize处理中,clock.c通过
servo_sample(c->servo, offset, tmv_to_nanoseconds(ingress), weight, &state)
servo->sample
函数中的算法计算得到当前Slave的4中状态
-
SERVO_UNLOCKED
-
servo尚未准备好跟踪Master
-
SERVO_JUMP
-
servo准备跟踪主时钟，但需要进行时钟跳跃来立即纠正估计的偏移
-
同时调用
clockadj_set_freq()和clockadj_step() -
SERVO_LOCKED
-
时钟同步已建立，正在进行正常的频率调整
-
clock_synchronize_locked()->clockadj_set_freq() -
SERVO_LOCKED_STABLE
-
伺服已稳定，最近的 'servo_num_offset_values' 个偏移估计值都小于 servo_offset_threshold
-
使用相位调整
clockadj_set_phase()
clockadj_step, clockadj_set_freq, clockadj_set_phase
分别是通过了系统apiclock_adjtime的ADJ_SETOFFSET ADJ_FREQUENCY ADJ_SETOFFSET进行时间调整
详细的参数使用,参考man/clock_adjtime
2.3.2 servo sample
servo 用于处理时钟偏移测量并计算出相应的频率调整值，以使从时钟能够同步到主时钟
LinuxPTP提供了多种servo实现
| PI Servo | pi.c | 最常用，比例-积分控制，适合大多数场景 |
| Linreg Servo | linreg.c | 线性回归，适合高精度需求 |
| NTP Servo | ntpshm.c | NTP算法，兼容NTP同步 |
| Null Servo | nullf.c | 空实现，测试或禁用servo |
可通过
servo配置项选择不同servo算法，具体实现可参考源码对应文件。
2.3.3 transport
LinuxPTP 支持2层报文和udp报文
初始化时通过参数来区分
// ptp4l.c
main()
{
case '2':
if (config_set_int(cfg, "network_transport",
TRANS_IEEE_802_3))
goto out;
break;
case '4':
if (config_set_int(cfg, "network_transport",
TRANS_UDP_IPV4))
goto out;
break;
case '6':
if (config_set_int(cfg, "network_transport",
TRANS_UDP_IPV6))
goto out;
break;
}
在创建transport时根据参数创建对应类型的port
// transport.c
struct transport *transport_create(struct config *cfg,
enum transport_type type)
{
struct transport *t = NULL;
switch (type) {
case TRANS_UDS:
t = uds_transport_create();
break;
case TRANS_UDP_IPV4:
t = udp_transport_create();
break;
case TRANS_UDP_IPV6:
t = udp6_transport_create();
break;
case TRANS_IEEE_802_3:
t = raw_transport_create();
最终udp_transport创建udp socket
// udp.c
udp_open
open_socket
fd = socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP);
raw_transport创建raw socket
// raw.c
raw_open
open_socket
fd = socket(PF_PACKET, SOCK_RAW, htons(ETH_P_ALL))
blog_link: japric.github.io/posts/linux…
